Bump ComfyUI version to v0.3.16

Fix bug when changing the compute dtype on wan.

README.md

@@ -1,7 +1,7 @@
 <div align="center">
 
 # ComfyUI
-**The most powerful and modular diffusion model GUI and backend.**
+**The most powerful and modular visual AI engine and application.**
 
 
 [![Website][website-shield]][website-url]
@@ -31,10 +31,24 @@
 ![ComfyUI Screenshot](https://github.com/user-attachments/assets/7ccaf2c1-9b72-41ae-9a89-5688c94b7abe)
 </div>
 
-This ui will let you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. For some workflow examples and see what ComfyUI can do you can check out:
-### [ComfyUI Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
+ComfyUI lets you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. Available on Windows, Linux, and macOS.
+
+## Get Started
+
+#### [Desktop Application](https://www.comfy.org/download)
+- The easiest way to get started. 
+- Available on Windows & macOS.
+
+#### [Windows Portable Package](#installing)
+- Get the latest commits and completely portable.
+- Available on Windows.
+
+#### [Manual Install](#manual-install-windows-linux)
+Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, Ascend).
+
+## Examples
+See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
 
-### [Installing ComfyUI](#installing)
 
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
@@ -121,7 +135,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 # Installing
 
-## Windows
+## Windows Portable
 
 There is a portable standalone build for Windows that should work for running on Nvidia GPUs or for running on your CPU only on the [releases page](https://github.com/comfyanonymous/ComfyUI/releases).
 
@@ -141,6 +155,15 @@ See the [Config file](extra_model_paths.yaml.example) to set the search paths fo
 
 To run it on services like paperspace, kaggle or colab you can use my [Jupyter Notebook](notebooks/comfyui_colab.ipynb)
 
+
+## [comfy-cli](https://docs.comfy.org/comfy-cli/getting-started)
+
+You can install and start ComfyUI using comfy-cli:
+```bash
+pip install comfy-cli
+comfy install
+```
+
 ## Manual Install (Windows, Linux)
 
 python 3.13 is supported but using 3.12 is recommended because some custom nodes and their dependencies might not support it yet.
@@ -293,6 +316,8 @@ Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app w
 
 ## Support and dev channel
 
+[Discord](https://comfy.org/discord): Try the #help or #feedback channels.
+
 [Matrix space: #comfyui_space:matrix.org](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) (it's like discord but open source).
 
 See also: [https://www.comfy.org/](https://www.comfy.org/)
@@ -309,7 +334,7 @@ For any bugs, issues, or feature requests related to the frontend, please use th
 
 The new frontend is now the default for ComfyUI. However, please note:
 
-1. The frontend in the main ComfyUI repository is updated weekly.
+1. The frontend in the main ComfyUI repository is updated fortnightly.
 2. Daily releases are available in the separate frontend repository.
 
 To use the most up-to-date frontend version:
@@ -326,7 +351,7 @@ To use the most up-to-date frontend version:
    --front-end-version Comfy-Org/ComfyUI_frontend@1.2.2
    ```
 
-This approach allows you to easily switch between the stable weekly release and the cutting-edge daily updates, or even specific versions for testing purposes.
+This approach allows you to easily switch between the stable fortnightly release and the cutting-edge daily updates, or even specific versions for testing purposes.
 
 ### Accessing the Legacy Frontend
 

api_server/routes/internal/internal_routes.py

@@ -1,9 +1,9 @@
 from aiohttp import web
 from typing import Optional
-from folder_paths import models_dir, user_directory, output_directory, folder_names_and_paths
-from api_server.services.file_service import FileService
+from folder_paths import folder_names_and_paths, get_directory_by_type
 from api_server.services.terminal_service import TerminalService
 import app.logger
+import os
 
 class InternalRoutes:
     '''
@@ -15,26 +15,10 @@ class InternalRoutes:
     def __init__(self, prompt_server):
         self.routes: web.RouteTableDef = web.RouteTableDef()
         self._app: Optional[web.Application] = None
-        self.file_service = FileService({
-            "models": models_dir,
-            "user": user_directory,
-            "output": output_directory
-        })
         self.prompt_server = prompt_server
         self.terminal_service = TerminalService(prompt_server)
 
     def setup_routes(self):
-        @self.routes.get('/files')
-        async def list_files(request):
-            directory_key = request.query.get('directory', '')
-            try:
-                file_list = self.file_service.list_files(directory_key)
-                return web.json_response({"files": file_list})
-            except ValueError as e:
-                return web.json_response({"error": str(e)}, status=400)
-            except Exception as e:
-                return web.json_response({"error": str(e)}, status=500)
-
         @self.routes.get('/logs')
         async def get_logs(request):
             return web.json_response("".join([(l["t"] + " - " + l["m"]) for l in app.logger.get_logs()]))
@@ -67,6 +51,20 @@ class InternalRoutes:
                 response[key] = folder_names_and_paths[key][0]
             return web.json_response(response)
 
+        @self.routes.get('/files/{directory_type}')
+        async def get_files(request: web.Request) -> web.Response:
+            directory_type = request.match_info['directory_type']
+            if directory_type not in ("output", "input", "temp"):
+                return web.json_response({"error": "Invalid directory type"}, status=400)
+
+            directory = get_directory_by_type(directory_type)
+            sorted_files = sorted(
+                (entry for entry in os.scandir(directory) if entry.is_file()),
+                key=lambda entry: -entry.stat().st_mtime
+            )
+            return web.json_response([entry.name for entry in sorted_files], status=200)
+
+
     def get_app(self):
         if self._app is None:
             self._app = web.Application()

api_server/services/file_service.py

@@ -1,13 +0,0 @@
-from typing import Dict, List, Optional
-from api_server.utils.file_operations import FileSystemOperations, FileSystemItem
-
-class FileService:
-    def __init__(self, allowed_directories: Dict[str, str], file_system_ops: Optional[FileSystemOperations] = None):
-        self.allowed_directories: Dict[str, str] = allowed_directories
-        self.file_system_ops: FileSystemOperations = file_system_ops or FileSystemOperations()
-
-    def list_files(self, directory_key: str) -> List[FileSystemItem]:
-        if directory_key not in self.allowed_directories:
-            raise ValueError("Invalid directory key")
-        directory_path: str = self.allowed_directories[directory_key]
-        return self.file_system_ops.walk_directory(directory_path)

comfy/cli_args.py

@@ -179,7 +179,7 @@ parser.add_argument(
 
 parser.add_argument("--user-directory", type=is_valid_directory, default=None, help="Set the ComfyUI user directory with an absolute path. Overrides --base-directory.")
 
-parser.add_argument("--disable-compres-response-body", action="store_true", help="Disable compressing response body.")
+parser.add_argument("--enable-compress-response-body", action="store_true", help="Enable compressing response body.")
 
 if comfy.options.args_parsing:
     args = parser.parse_args()
@@ -191,3 +191,6 @@ if args.windows_standalone_build:
 
 if args.disable_auto_launch:
     args.auto_launch = False
+
+if args.force_fp16:
+    args.fp16_unet = True

comfy/clip_model.py

@@ -104,7 +104,8 @@ class CLIPTextModel_(torch.nn.Module):
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
             mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max)
 
-        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(-torch.finfo(x.dtype).max).triu_(1)
+        causal_mask = torch.full((x.shape[1], x.shape[1]), -torch.finfo(x.dtype).max, dtype=x.dtype, device=x.device).triu_(1)
+
         if mask is not None:
             mask += causal_mask
         else:

comfy/comfy_types/node_typing.py

@@ -66,13 +66,26 @@ class IO(StrEnum):
         b = frozenset(value.split(","))
         return not (b.issubset(a) or a.issubset(b))
 
+class RemoteInputOptions(TypedDict):
+    route: str
+    """The route to the remote source."""
+    refresh_button: bool
+    """Specifies whether to show a refresh button in the UI below the widget."""
+    control_after_refresh: Literal["first", "last"]
+    """Specifies the control after the refresh button is clicked. If "first", the first item will be automatically selected, and so on."""
+    timeout: int
+    """The maximum amount of time to wait for a response from the remote source in milliseconds."""
+    max_retries: int
+    """The maximum number of retries before aborting the request."""
+    refresh: int
+    """The TTL of the remote input's value in milliseconds. Specifies the interval at which the remote input's value is refreshed."""
 
 class InputTypeOptions(TypedDict):
     """Provides type hinting for the return type of the INPUT_TYPES node function.
 
     Due to IDE limitations with unions, for now all options are available for all types (e.g. `label_on` is hinted even when the type is not `IO.BOOLEAN`).
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_datatypes
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/datatypes
     """
 
     default: bool | str | float | int | list | tuple
@@ -113,6 +126,14 @@ class InputTypeOptions(TypedDict):
     # defaultVal: str
     dynamicPrompts: bool
     """Causes the front-end to evaluate dynamic prompts (``STRING``)"""
+    # class InputTypeCombo(InputTypeOptions):
+    image_upload: bool
+    """Specifies whether the input should have an image upload button and image preview attached to it. Requires that the input's name is `image`."""
+    image_folder: Literal["input", "output", "temp"]
+    """Specifies which folder to get preview images from if the input has the ``image_upload`` flag.
+    """
+    remote: RemoteInputOptions
+    """Specifies the configuration for a remote input."""
 
 
 class HiddenInputTypeDict(TypedDict):
@@ -133,7 +154,7 @@ class HiddenInputTypeDict(TypedDict):
 class InputTypeDict(TypedDict):
     """Provides type hinting for node INPUT_TYPES.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_more_on_inputs
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/more_on_inputs
     """
 
     required: dict[str, tuple[IO, InputTypeOptions]]
@@ -143,14 +164,14 @@ class InputTypeDict(TypedDict):
     hidden: HiddenInputTypeDict
     """Offers advanced functionality and server-client communication.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_more_on_inputs#hidden-inputs
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/more_on_inputs#hidden-inputs
     """
 
 
 class ComfyNodeABC(ABC):
     """Abstract base class for Comfy nodes.  Includes the names and expected types of attributes.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview
     """
 
     DESCRIPTION: str
@@ -167,7 +188,7 @@ class ComfyNodeABC(ABC):
     CATEGORY: str
     """The category of the node, as per the "Add Node" menu.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#category
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#category
     """
     EXPERIMENTAL: bool
     """Flags a node as experimental, informing users that it may change or not work as expected."""
@@ -181,9 +202,9 @@ class ComfyNodeABC(ABC):
 
         * Must include the ``required`` key, which describes all inputs that must be connected for the node to execute.
         * The ``optional`` key can be added to describe inputs which do not need to be connected.
-        * The ``hidden`` key offers some advanced functionality.  More info at: https://docs.comfy.org/essentials/custom_node_more_on_inputs#hidden-inputs
+        * The ``hidden`` key offers some advanced functionality.  More info at: https://docs.comfy.org/custom-nodes/backend/more_on_inputs#hidden-inputs
 
-        Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#input-types
+        Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#input-types
         """
         return {"required": {}}
 
@@ -198,7 +219,7 @@ class ComfyNodeABC(ABC):
 
     By default, a node is not considered an output. Set ``OUTPUT_NODE = True`` to specify that it is.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#output-node
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#output-node
     """
     INPUT_IS_LIST: bool
     """A flag indicating if this node implements the additional code necessary to deal with OUTPUT_IS_LIST nodes.
@@ -209,7 +230,7 @@ class ComfyNodeABC(ABC):
 
     A node can also override the default input behaviour and receive the whole list in a single call. This is done by setting a class attribute `INPUT_IS_LIST` to ``True``.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_lists#list-processing
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
     OUTPUT_IS_LIST: tuple[bool]
     """A tuple indicating which node outputs are lists, but will be connected to nodes that expect individual items.
@@ -227,7 +248,7 @@ class ComfyNodeABC(ABC):
     the node should provide a class attribute `OUTPUT_IS_LIST`, which is a ``tuple[bool]``, of the same length as `RETURN_TYPES`,
     specifying which outputs which should be so treated.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_lists#list-processing
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
 
     RETURN_TYPES: tuple[IO]
@@ -237,19 +258,19 @@ class ComfyNodeABC(ABC):
 
         RETURN_TYPES = (IO.INT, "INT", "CUSTOM_TYPE")
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#return-types
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-types
     """
     RETURN_NAMES: tuple[str]
     """The output slot names for each item in `RETURN_TYPES`, e.g. ``RETURN_NAMES = ("count", "filter_string")``
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#return-names
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-names
     """
     OUTPUT_TOOLTIPS: tuple[str]
     """A tuple of strings to use as tooltips for node outputs, one for each item in `RETURN_TYPES`."""
     FUNCTION: str
     """The name of the function to execute as a literal string, e.g. `FUNCTION = "execute"`
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#function
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#function
     """
 
 
@@ -267,7 +288,7 @@ class CheckLazyMixin:
         Params should match the nodes execution ``FUNCTION`` (self, and all inputs by name).
         Will be executed repeatedly until it returns an empty list, or all requested items were already evaluated (and sent as params).
 
-        Comfy Docs: https://docs.comfy.org/essentials/custom_node_lazy_evaluation#defining-check-lazy-status
+        Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lazy_evaluation#defining-check-lazy-status
         """
 
         need = [name for name in kwargs if kwargs[name] is None]

comfy/k_diffusion/sampling.py

@@ -1267,7 +1267,7 @@ def sample_dpmpp_2m_cfg_pp(model, x, sigmas, extra_args=None, callback=None, dis
     return x
 
 @torch.no_grad()
-def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1., noise_sampler=None, cfg_pp=False):
+def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, eta=1., cfg_pp=False):
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
@@ -1289,53 +1289,60 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
         extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
 
     for i in trange(len(sigmas) - 1, disable=disable):
-        if s_churn > 0:
-            gamma = min(s_churn / (len(sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.0
-            sigma_hat = sigmas[i] * (gamma + 1)
-        else:
-            gamma = 0
-            sigma_hat = sigmas[i]
-
-        if gamma > 0:
-            eps = torch.randn_like(x) * s_noise
-            x = x + eps * (sigma_hat**2 - sigmas[i] ** 2) ** 0.5
-        denoised = model(x, sigma_hat * s_in, **extra_args)
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
         if callback is not None:
-            callback({"x": x, "i": i, "sigma": sigmas[i], "sigma_hat": sigma_hat, "denoised": denoised})
-        if sigmas[i + 1] == 0 or old_denoised is None:
+            callback({"x": x, "i": i, "sigma": sigmas[i], "sigma_hat": sigmas[i], "denoised": denoised})
+        if sigma_down == 0 or old_denoised is None:
             # Euler method
             if cfg_pp:
-                d = to_d(x, sigma_hat, uncond_denoised)
-                x = denoised + d * sigmas[i + 1]
+                d = to_d(x, sigmas[i], uncond_denoised)
+                x = denoised + d * sigma_down
             else:
-                d = to_d(x, sigma_hat, denoised)
-                dt = sigmas[i + 1] - sigma_hat
+                d = to_d(x, sigmas[i], denoised)
+                dt = sigma_down - sigmas[i]
                 x = x + d * dt
         else:
             # Second order multistep method in https://arxiv.org/pdf/2308.02157
-            t, t_next, t_prev = t_fn(sigmas[i]), t_fn(sigmas[i + 1]), t_fn(sigmas[i - 1])
+            t, t_next, t_prev = t_fn(sigmas[i]), t_fn(sigma_down), t_fn(sigmas[i - 1])
             h = t_next - t
             c2 = (t_prev - t) / h
 
             phi1_val, phi2_val = phi1_fn(-h), phi2_fn(-h)
-            b1 = torch.nan_to_num(phi1_val - 1.0 / c2 * phi2_val, nan=0.0)
-            b2 = torch.nan_to_num(1.0 / c2 * phi2_val, nan=0.0)
+            b1 = torch.nan_to_num(phi1_val - phi2_val / c2, nan=0.0)
+            b2 = torch.nan_to_num(phi2_val / c2, nan=0.0)
 
             if cfg_pp:
                 x = x + (denoised - uncond_denoised)
+                x = sigma_fn(h) * x + h * (b1 * uncond_denoised + b2 * old_denoised)
+            else:
+                x = sigma_fn(h) * x + h * (b1 * denoised + b2 * old_denoised)
 
-            x = (sigma_fn(t_next) / sigma_fn(t)) * x + h * (b1 * denoised + b2 * old_denoised)
+        # Noise addition
+        if sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
 
-        old_denoised = denoised
+        if cfg_pp:
+            old_denoised = uncond_denoised
+        else:
+            old_denoised = denoised
     return x
 
 @torch.no_grad()
-def sample_res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1., noise_sampler=None):
-    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_churn=s_churn, s_tmin=s_tmin, s_tmax=s_tmax, s_noise=s_noise, noise_sampler=noise_sampler, cfg_pp=False)
+def sample_res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=0., cfg_pp=False)
 
 @torch.no_grad()
-def sample_res_multistep_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1., noise_sampler=None):
-    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_churn=s_churn, s_tmin=s_tmin, s_tmax=s_tmax, s_noise=s_noise, noise_sampler=noise_sampler, cfg_pp=True)
+def sample_res_multistep_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=0., cfg_pp=True)
+
+@torch.no_grad()
+def sample_res_multistep_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=False)
+
+@torch.no_grad()
+def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True)
 
 @torch.no_grad()
 def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):

comfy/latent_formats.py

@@ -407,3 +407,52 @@ class Cosmos1CV8x8x8(LatentFormat):
     ]
 
     latent_rgb_factors_bias = [-0.1223, -0.1889, -0.1976]
+
+class Wan21(LatentFormat):
+    latent_channels = 16
+    latent_dimensions = 3
+
+    latent_rgb_factors = [
+            [-0.1299, -0.1692,  0.2932],
+            [ 0.0671,  0.0406,  0.0442],
+            [ 0.3568,  0.2548,  0.1747],
+            [ 0.0372,  0.2344,  0.1420],
+            [ 0.0313,  0.0189, -0.0328],
+            [ 0.0296, -0.0956, -0.0665],
+            [-0.3477, -0.4059, -0.2925],
+            [ 0.0166,  0.1902,  0.1975],
+            [-0.0412,  0.0267, -0.1364],
+            [-0.1293,  0.0740,  0.1636],
+            [ 0.0680,  0.3019,  0.1128],
+            [ 0.0032,  0.0581,  0.0639],
+            [-0.1251,  0.0927,  0.1699],
+            [ 0.0060, -0.0633,  0.0005],
+            [ 0.3477,  0.2275,  0.2950],
+            [ 0.1984,  0.0913,  0.1861]
+        ]
+
+    latent_rgb_factors_bias = [-0.1835, -0.0868, -0.3360]
+
+    def __init__(self):
+        self.scale_factor = 1.0
+        self.latents_mean = torch.tensor([
+            -0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508,
+            0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921
+        ]).view(1, self.latent_channels, 1, 1, 1)
+        self.latents_std = torch.tensor([
+            2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743,
+            3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160
+        ]).view(1, self.latent_channels, 1, 1, 1)
+
+
+        self.taesd_decoder_name = None #TODO
+
+    def process_in(self, latent):
+        latents_mean = self.latents_mean.to(latent.device, latent.dtype)
+        latents_std = self.latents_std.to(latent.device, latent.dtype)
+        return (latent - latents_mean) * self.scale_factor / latents_std
+
+    def process_out(self, latent):
+        latents_mean = self.latents_mean.to(latent.device, latent.dtype)
+        latents_std = self.latents_std.to(latent.device, latent.dtype)
+        return latent * latents_std / self.scale_factor + latents_mean

comfy/ldm/flux/math.py

@@ -22,7 +22,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu():
+    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
         device = torch.device("cpu")
     else:
         device = pos.device

comfy/ldm/hunyuan_video/model.py

@@ -310,7 +310,7 @@ class HunyuanVideo(nn.Module):
             shape[i] = shape[i] // self.patch_size[i]
         img = img.reshape([img.shape[0]] + shape + [self.out_channels] + self.patch_size)
         img = img.permute(0, 4, 1, 5, 2, 6, 3, 7)
-        img = img.reshape(initial_shape)
+        img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4])
         return img
 
     def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, control=None, transformer_options={}, **kwargs):

comfy/ldm/lumina/model.py

@@ -6,6 +6,7 @@ from typing import List, Optional, Tuple
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+import comfy.ldm.common_dit
 
 from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, RMSNorm
 from comfy.ldm.modules.attention import optimized_attention_masked
@@ -352,25 +353,6 @@ class FinalLayer(nn.Module):
         return x
 
 
-class RopeEmbedder:
-    def __init__(
-        self, theta: float = 10000.0, axes_dims: List[int] = (16, 56, 56), axes_lens: List[int] = (1, 512, 512)
-    ):
-        super().__init__()
-        self.theta = theta
-        self.axes_dims = axes_dims
-        self.axes_lens = axes_lens
-        self.freqs_cis = NextDiT.precompute_freqs_cis(self.axes_dims, self.axes_lens, theta=self.theta)
-
-    def __call__(self, ids: torch.Tensor):
-        self.freqs_cis = [freqs_cis.to(ids.device) for freqs_cis in self.freqs_cis]
-        result = []
-        for i in range(len(self.axes_dims)):
-            index = ids[:, :, i:i+1].repeat(1, 1, self.freqs_cis[i].shape[-1]).to(torch.int64)
-            result.append(torch.gather(self.freqs_cis[i].unsqueeze(0).repeat(index.shape[0], 1, 1), dim=1, index=index))
-        return torch.cat(result, dim=-1)
-
-
 class NextDiT(nn.Module):
     """
     Diffusion model with a Transformer backbone.
@@ -481,7 +463,6 @@ class NextDiT(nn.Module):
         assert (dim // n_heads) == sum(axes_dims)
         self.axes_dims = axes_dims
         self.axes_lens = axes_lens
-        # self.rope_embedder = RopeEmbedder(axes_dims=axes_dims, axes_lens=axes_lens)
         self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims)
         self.dim = dim
         self.n_heads = n_heads
@@ -609,12 +590,13 @@ class NextDiT(nn.Module):
 
         return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
 
-
     # def forward(self, x, t, cap_feats, cap_mask):
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask
+        bs, c, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
         """
         Forward pass of NextDiT.
         t: (N,) tensor of diffusion timesteps
@@ -634,41 +616,7 @@ class NextDiT(nn.Module):
             x = layer(x, mask, freqs_cis, adaln_input)
 
         x = self.final_layer(x, adaln_input)
-        x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)
+        x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]
 
         return -x
 
-    @staticmethod
-    def precompute_freqs_cis(
-        dim: List[int],
-        end: List[int],
-        theta: float = 10000.0,
-    ):
-        """
-        Precompute the frequency tensor for complex exponentials (cis) with
-        given dimensions.
-
-        This function calculates a frequency tensor with complex exponentials
-        using the given dimension 'dim' and the end index 'end'. The 'theta'
-        parameter scales the frequencies. The returned tensor contains complex
-        values in complex64 data type.
-
-        Args:
-            dim (list): Dimension of the frequency tensor.
-            end (list): End index for precomputing frequencies.
-            theta (float, optional): Scaling factor for frequency computation.
-                Defaults to 10000.0.
-
-        Returns:
-            torch.Tensor: Precomputed frequency tensor with complex
-                exponentials.
-        """
-        freqs_cis = []
-        for i, (d, e) in enumerate(zip(dim, end)):
-            freqs = 1.0 / (theta ** (torch.arange(0, d, 2, dtype=torch.float64, device="cpu") / d))
-            timestep = torch.arange(e, device=freqs.device, dtype=torch.float64)
-            freqs = torch.outer(timestep, freqs).float()
-            freqs_cis_i = torch.polar(torch.ones_like(freqs), freqs).to(torch.complex64)  # complex64
-            freqs_cis.append(freqs_cis_i)
-
-        return freqs_cis

comfy/ldm/modules/attention.py

@@ -30,38 +30,24 @@ ops = comfy.ops.disable_weight_init
 
 FORCE_UPCAST_ATTENTION_DTYPE = model_management.force_upcast_attention_dtype()
 
-def get_attn_precision(attn_precision):
+def get_attn_precision(attn_precision, current_dtype):
     if args.dont_upcast_attention:
         return None
-    if FORCE_UPCAST_ATTENTION_DTYPE is not None:
-        return FORCE_UPCAST_ATTENTION_DTYPE
+
+    if FORCE_UPCAST_ATTENTION_DTYPE is not None and current_dtype in FORCE_UPCAST_ATTENTION_DTYPE:
+        return FORCE_UPCAST_ATTENTION_DTYPE[current_dtype]
     return attn_precision
 
 def exists(val):
     return val is not None
 
 
-def uniq(arr):
-    return{el: True for el in arr}.keys()
-
-
 def default(val, d):
     if exists(val):
         return val
     return d
 
 
-def max_neg_value(t):
-    return -torch.finfo(t.dtype).max
-
-
-def init_(tensor):
-    dim = tensor.shape[-1]
-    std = 1 / math.sqrt(dim)
-    tensor.uniform_(-std, std)
-    return tensor
-
-
 # feedforward
 class GEGLU(nn.Module):
     def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=ops):
@@ -96,7 +82,7 @@ def Normalize(in_channels, dtype=None, device=None):
     return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
 
 def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = q.shape
@@ -165,7 +151,7 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
 
 
 def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, query.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = query.shape
@@ -235,7 +221,7 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
     return hidden_states
 
 def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = q.shape

comfy/ldm/modules/diffusionmodules/model.py

@@ -297,7 +297,7 @@ def vae_attention():
     if model_management.xformers_enabled_vae():
         logging.info("Using xformers attention in VAE")
         return xformers_attention
-    elif model_management.pytorch_attention_enabled():
+    elif model_management.pytorch_attention_enabled_vae():
         logging.info("Using pytorch attention in VAE")
         return pytorch_attention
     else:

comfy/ldm/wan/model.py

@@ -0,0 +1,485 @@
+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/model.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import math
+
+import torch
+import torch.nn as nn
+from einops import repeat
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
+from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm
+import comfy.ldm.common_dit
+import comfy.model_management
+
+
+def sinusoidal_embedding_1d(dim, position):
+    # preprocess
+    assert dim % 2 == 0
+    half = dim // 2
+    position = position.type(torch.float64)
+
+    # calculation
+    sinusoid = torch.outer(
+        position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x
+
+
+class WanSelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+
+        # layers
+        self.q = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.k = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.o = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.norm_q = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_k = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+
+    def forward(self, x, freqs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, num_heads, C / num_heads]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        # query, key, value function
+        def qkv_fn(x):
+            q = self.norm_q(self.q(x)).view(b, s, n, d)
+            k = self.norm_k(self.k(x)).view(b, s, n, d)
+            v = self.v(x).view(b, s, n * d)
+            return q, k, v
+
+        q, k, v = qkv_fn(x)
+        q, k = apply_rope(q, k, freqs)
+
+        x = optimized_attention(
+            q.view(b, s, n * d),
+            k.view(b, s, n * d),
+            v,
+            heads=self.num_heads,
+        )
+
+        x = self.o(x)
+        return x
+
+
+class WanT2VCrossAttention(WanSelfAttention):
+
+    def forward(self, x, context):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        x = self.o(x)
+        return x
+
+
+class WanI2VCrossAttention(WanSelfAttention):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        super().__init__(dim, num_heads, window_size, qk_norm, eps, operation_settings=operation_settings)
+
+        self.k_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.alpha = nn.Parameter(torch.zeros((1, )))
+        self.norm_k_img = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+
+    def forward(self, x, context):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        context_img = context[:, :257]
+        context = context[:, 257:]
+
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+        k_img = self.norm_k_img(self.k_img(context_img))
+        v_img = self.v_img(context_img)
+        img_x = optimized_attention(q, k_img, v_img, heads=self.num_heads)
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        # output
+        x = x + img_x
+        x = self.o(x)
+        return x
+
+
+WAN_CROSSATTENTION_CLASSES = {
+    't2v_cross_attn': WanT2VCrossAttention,
+    'i2v_cross_attn': WanI2VCrossAttention,
+}
+
+
+class WanAttentionBlock(nn.Module):
+
+    def __init__(self,
+                 cross_attn_type,
+                 dim,
+                 ffn_dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=False,
+                 eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # layers
+        self.norm1 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm,
+                                          eps, operation_settings=operation_settings)
+        self.norm3 = operation_settings.get("operations").LayerNorm(
+            dim, eps,
+            elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if cross_attn_norm else nn.Identity()
+        self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim,
+                                                                      num_heads,
+                                                                      (-1, -1),
+                                                                      qk_norm,
+                                                                      eps, operation_settings=operation_settings)
+        self.norm2 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn = nn.Sequential(
+            operation_settings.get("operations").Linear(dim, ffn_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operation_settings.get("operations").Linear(ffn_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 6, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(
+        self,
+        x,
+        e,
+        freqs,
+        context,
+    ):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, C]
+            e(Tensor): Shape [B, 6, C]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        # assert e.dtype == torch.float32
+
+        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        # assert e[0].dtype == torch.float32
+
+        # self-attention
+        y = self.self_attn(
+            self.norm1(x) * (1 + e[1]) + e[0],
+            freqs)
+
+        x = x + y * e[2]
+
+        # cross-attention & ffn function
+        def cross_attn_ffn(x, context, e):
+            x = x + self.cross_attn(self.norm3(x), context)
+            y = self.ffn(self.norm2(x) * (1 + e[4]) + e[3])
+            x = x + y * e[5]
+            return x
+
+        x = cross_attn_ffn(x, context, e)
+        return x
+
+
+class Head(nn.Module):
+
+    def __init__(self, dim, out_dim, patch_size, eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.eps = eps
+
+        # layers
+        out_dim = math.prod(patch_size) * out_dim
+        self.norm = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.head = operation_settings.get("operations").Linear(dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 2, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(self, x, e):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            e(Tensor): Shape [B, C]
+        """
+        # assert e.dtype == torch.float32
+        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
+        x = (self.head(self.norm(x) * (1 + e[1]) + e[0]))
+        return x
+
+
+class MLPProj(torch.nn.Module):
+
+    def __init__(self, in_dim, out_dim, operation_settings={}):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            operation_settings.get("operations").LayerNorm(in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), operation_settings.get("operations").Linear(in_dim, in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            torch.nn.GELU(), operation_settings.get("operations").Linear(in_dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            operation_settings.get("operations").LayerNorm(out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class WanModel(torch.nn.Module):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='t2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        r"""
+        Initialize the diffusion model backbone.
+
+        Args:
+            model_type (`str`, *optional*, defaults to 't2v'):
+                Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video)
+            patch_size (`tuple`, *optional*, defaults to (1, 2, 2)):
+                3D patch dimensions for video embedding (t_patch, h_patch, w_patch)
+            text_len (`int`, *optional*, defaults to 512):
+                Fixed length for text embeddings
+            in_dim (`int`, *optional*, defaults to 16):
+                Input video channels (C_in)
+            dim (`int`, *optional*, defaults to 2048):
+                Hidden dimension of the transformer
+            ffn_dim (`int`, *optional*, defaults to 8192):
+                Intermediate dimension in feed-forward network
+            freq_dim (`int`, *optional*, defaults to 256):
+                Dimension for sinusoidal time embeddings
+            text_dim (`int`, *optional*, defaults to 4096):
+                Input dimension for text embeddings
+            out_dim (`int`, *optional*, defaults to 16):
+                Output video channels (C_out)
+            num_heads (`int`, *optional*, defaults to 16):
+                Number of attention heads
+            num_layers (`int`, *optional*, defaults to 32):
+                Number of transformer blocks
+            window_size (`tuple`, *optional*, defaults to (-1, -1)):
+                Window size for local attention (-1 indicates global attention)
+            qk_norm (`bool`, *optional*, defaults to True):
+                Enable query/key normalization
+            cross_attn_norm (`bool`, *optional*, defaults to False):
+                Enable cross-attention normalization
+            eps (`float`, *optional*, defaults to 1e-6):
+                Epsilon value for normalization layers
+        """
+
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        assert model_type in ['t2v', 'i2v']
+        self.model_type = model_type
+
+        self.patch_size = patch_size
+        self.text_len = text_len
+        self.in_dim = in_dim
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.freq_dim = freq_dim
+        self.text_dim = text_dim
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # embeddings
+        self.patch_embedding = operations.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.text_embedding = nn.Sequential(
+            operations.Linear(text_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        self.time_embedding = nn.Sequential(
+            operations.Linear(freq_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.SiLU(), operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+        self.time_projection = nn.Sequential(nn.SiLU(), operations.Linear(dim, dim * 6, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # blocks
+        cross_attn_type = 't2v_cross_attn' if model_type == 't2v' else 'i2v_cross_attn'
+        self.blocks = nn.ModuleList([
+            WanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads,
+                              window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
+            for _ in range(num_layers)
+        ])
+
+        # head
+        self.head = Head(dim, out_dim, patch_size, eps, operation_settings=operation_settings)
+
+        d = dim // num_heads
+        self.rope_embedder = EmbedND(dim=d, theta=10000.0, axes_dim=[d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)])
+
+        if model_type == 'i2v':
+            self.img_emb = MLPProj(1280, dim, operation_settings=operation_settings)
+        else:
+            self.img_emb = None
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+    ):
+        r"""
+        Forward pass through the diffusion model
+
+        Args:
+            x (Tensor):
+                List of input video tensors with shape [B, C_in, F, H, W]
+            t (Tensor):
+                Diffusion timesteps tensor of shape [B]
+            context (List[Tensor]):
+                List of text embeddings each with shape [B, L, C]
+            seq_len (`int`):
+                Maximum sequence length for positional encoding
+            clip_fea (Tensor, *optional*):
+                CLIP image features for image-to-video mode
+            y (List[Tensor], *optional*):
+                Conditional video inputs for image-to-video mode, same shape as x
+
+        Returns:
+            List[Tensor]:
+                List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
+        """
+        # embeddings
+        x = self.patch_embedding(x)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
+        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+
+        # context
+        context = self.text_embedding(torch.cat([context, context.new_zeros(context.size(0), self.text_len - context.size(1), context.size(2))], dim=1))
+
+        if clip_fea is not None and self.img_emb is not None:
+            context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+            context = torch.concat([context_clip, context], dim=1)
+
+        # arguments
+        kwargs = dict(
+            e=e0,
+            freqs=freqs,
+            context=context)
+
+        for block in self.blocks:
+            x = block(x, **kwargs)
+
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+        # return [u.float() for u in x]
+
+    def forward(self, x, timestep, context, clip_fea=None, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
+        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs)[:, :, :t, :h, :w]
+
+    def unpatchify(self, x, grid_sizes):
+        r"""
+        Reconstruct video tensors from patch embeddings.
+
+        Args:
+            x (List[Tensor]):
+                List of patchified features, each with shape [L, C_out * prod(patch_size)]
+            grid_sizes (Tensor):
+                Original spatial-temporal grid dimensions before patching,
+                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
+
+        Returns:
+            List[Tensor]:
+                Reconstructed video tensors with shape [L, C_out, F, H / 8, W / 8]
+        """
+
+        c = self.out_dim
+        u = x
+        b = u.shape[0]
+        u = u[:, :math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c)
+        u = torch.einsum('bfhwpqrc->bcfphqwr', u)
+        u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)])
+        return u

comfy/ldm/wan/vae.py

@@ -0,0 +1,567 @@
+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/vae.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from comfy.ldm.modules.diffusionmodules.model import vae_attention
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class CausalConv3d(ops.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else None
+
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0)
+
+
+class Upsample(nn.Upsample):
+
+    def forward(self, x):
+        """
+        Fix bfloat16 support for nearest neighbor interpolation.
+        """
+        return super().forward(x.float()).type_as(x)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                                            dim=2)
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    # if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx]!='Rep':
+                    #     # cache last frame of last two chunk
+                    #     cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+    def init_weight(self, conv):
+        conv_weight = conv.weight
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        one_matrix = torch.eye(c1, c2)
+        init_matrix = one_matrix
+        nn.init.zeros_(conv_weight)
+        #conv_weight.data[:,:,-1,1,1] = init_matrix * 0.5
+        conv_weight.data[:, :, 1, 0, 0] = init_matrix  #* 0.5
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def init_weight2(self, conv):
+        conv_weight = conv.weight.data
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        init_matrix = torch.eye(c1 // 2, c2)
+        #init_matrix = repeat(init_matrix, 'o ... -> (o 2) ...').permute(1,0,2).contiguous().reshape(c1,c2)
+        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
+        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        h = self.shortcut(x)
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = ops.Conv2d(dim, dim * 3, 1)
+        self.proj = ops.Conv2d(dim, dim, 1)
+        self.optimized_attention = vae_attention()
+
+    def forward(self, x):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+
+        q, k, v = self.to_qkv(x).chunk(3, dim=1)
+        x = self.optimized_attention(q, k, v)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout), AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout), AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, 3, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def forward(self, x):
+        mu, log_var = self.encode(x)
+        z = self.reparameterize(mu, log_var)
+        x_recon = self.decode(z)
+        return x_recon, mu, log_var
+
+    def encode(self, x):
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        ## 对encode输入的x，按时间拆分为1、4、4、4....
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
+        return mu
+
+    def decode(self, z):
+        self.clear_cache()
+        # z: [b,c,t,h,w]
+
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+        self.clear_cache()
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        #cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/model_base.py

@@ -35,6 +35,7 @@ import comfy.ldm.lightricks.model
 import comfy.ldm.hunyuan_video.model
 import comfy.ldm.cosmos.model
 import comfy.ldm.lumina.model
+import comfy.ldm.wan.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -166,9 +167,6 @@ class BaseModel(torch.nn.Module):
     def get_dtype(self):
         return self.diffusion_model.dtype
 
-    def is_adm(self):
-        return self.adm_channels > 0
-
     def encode_adm(self, **kwargs):
         return None
 
@@ -874,6 +872,15 @@ class HunyuanVideo(BaseModel):
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
 
+        image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+
+        if image is not None:
+            padding_shape = (noise.shape[0], 16, noise.shape[2] - 1, noise.shape[3], noise.shape[4])
+            latent_padding = torch.zeros(padding_shape, device=noise.device, dtype=noise.dtype)
+            image_latents = torch.cat([image.to(noise), latent_padding], dim=2)
+            out['c_concat'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_latents))
+
         guidance = kwargs.get("guidance", 6.0)
         if guidance is not None:
             out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
@@ -921,3 +928,47 @@ class Lumina2(BaseModel):
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
+
+class WAN21(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
+    def concat_cond(self, **kwargs):
+        if not self.image_to_video:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+        device = kwargs["device"]
+
+        if image is None:
+            image = torch.zeros_like(noise)
+
+        image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+        image = self.process_latent_in(image)
+        image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :4]
+        else:
+            mask = 1.0 - torch.mean(mask, dim=1, keepdim=True)
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            mask = mask.repeat(1, 4, 1, 1, 1)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        return torch.cat((mask, image), dim=1)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        clip_vision_output = kwargs.get("clip_vision_output", None)
+        if clip_vision_output is not None:
+            out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
+        return out

comfy/model_detection.py

@@ -136,7 +136,7 @@ def detect_unet_config(state_dict, key_prefix):
     if '{}txt_in.individual_token_refiner.blocks.0.norm1.weight'.format(key_prefix) in state_dict_keys: #Hunyuan Video
         dit_config = {}
         dit_config["image_model"] = "hunyuan_video"
-        dit_config["in_channels"] = 16
+        dit_config["in_channels"] = state_dict['{}img_in.proj.weight'.format(key_prefix)].shape[1] #SkyReels img2video has 32 input channels
         dit_config["patch_size"] = [1, 2, 2]
         dit_config["out_channels"] = 16
         dit_config["vec_in_dim"] = 768
@@ -299,6 +299,27 @@ def detect_unet_config(state_dict, key_prefix):
         dit_config["axes_lens"] = [300, 512, 512]
         return dit_config
 
+    if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
+        dit_config = {}
+        dit_config["image_model"] = "wan2.1"
+        dim = state_dict['{}head.modulation'.format(key_prefix)].shape[-1]
+        dit_config["dim"] = dim
+        dit_config["num_heads"] = dim // 128
+        dit_config["ffn_dim"] = state_dict['{}blocks.0.ffn.0.weight'.format(key_prefix)].shape[0]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+        dit_config["patch_size"] = (1, 2, 2)
+        dit_config["freq_dim"] = 256
+        dit_config["window_size"] = (-1, -1)
+        dit_config["qk_norm"] = True
+        dit_config["cross_attn_norm"] = True
+        dit_config["eps"] = 1e-6
+        dit_config["in_dim"] = state_dict['{}patch_embedding.weight'.format(key_prefix)].shape[1]
+        if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "i2v"
+        else:
+            dit_config["model_type"] = "t2v"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/model_management.py

@@ -50,7 +50,9 @@ xpu_available = False
 torch_version = ""
 try:
     torch_version = torch.version.__version__
-    xpu_available = (int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) <= 4)) and torch.xpu.is_available()
+    temp = torch_version.split(".")
+    torch_version_numeric = (int(temp[0]), int(temp[1]))
+    xpu_available = (torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] <= 4)) and torch.xpu.is_available()
 except:
     pass
 
@@ -218,7 +220,7 @@ def is_amd():
 
 MIN_WEIGHT_MEMORY_RATIO = 0.4
 if is_nvidia():
-    MIN_WEIGHT_MEMORY_RATIO = 0.1
+    MIN_WEIGHT_MEMORY_RATIO = 0.0
 
 ENABLE_PYTORCH_ATTENTION = False
 if args.use_pytorch_cross_attention:
@@ -227,7 +229,7 @@ if args.use_pytorch_cross_attention:
 
 try:
     if is_nvidia():
-        if int(torch_version[0]) >= 2:
+        if torch_version_numeric[0] >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
     if is_intel_xpu() or is_ascend_npu():
@@ -236,13 +238,35 @@ try:
 except:
     pass
 
+
+try:
+    if is_amd():
+        arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
+        logging.info("AMD arch: {}".format(arch))
+        if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
+            if torch_version_numeric[0] >= 2 and torch_version_numeric[1] >= 7:  # works on 2.6 but doesn't actually seem to improve much
+                if any((a in arch) for a in ["gfx1100", "gfx1101"]):  # TODO: more arches
+                    ENABLE_PYTORCH_ATTENTION = True
+except:
+    pass
+
+
 if ENABLE_PYTORCH_ATTENTION:
     torch.backends.cuda.enable_math_sdp(True)
     torch.backends.cuda.enable_flash_sdp(True)
     torch.backends.cuda.enable_mem_efficient_sdp(True)
 
+
+PRIORITIZE_FP16 = False  # TODO: remove and replace with something that shows exactly which dtype is faster than the other
 try:
-    if int(torch_version[0]) == 2 and int(torch_version[2]) >= 5:
+    if is_nvidia() and args.fast:
+        torch.backends.cuda.matmul.allow_fp16_accumulation = True
+        PRIORITIZE_FP16 = True  # TODO: limit to cards where it actually boosts performance
+except:
+    pass
+
+try:
+    if torch_version_numeric[0] == 2 and torch_version_numeric[1] >= 5:
         torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)
 except:
     logging.warning("Warning, could not set allow_fp16_bf16_reduction_math_sdp")
@@ -256,15 +280,10 @@ elif args.highvram or args.gpu_only:
     vram_state = VRAMState.HIGH_VRAM
 
 FORCE_FP32 = False
-FORCE_FP16 = False
 if args.force_fp32:
     logging.info("Forcing FP32, if this improves things please report it.")
     FORCE_FP32 = True
 
-if args.force_fp16:
-    logging.info("Forcing FP16.")
-    FORCE_FP16 = True
-
 if lowvram_available:
     if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
         vram_state = set_vram_to
@@ -665,6 +684,10 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         if model_params * 2 > free_model_memory:
             return fp8_dtype
 
+    if PRIORITIZE_FP16:
+        if torch.float16 in supported_dtypes and should_use_fp16(device=device, model_params=model_params):
+            return torch.float16
+
     for dt in supported_dtypes:
         if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params):
             if torch.float16 in supported_dtypes:
@@ -898,6 +921,11 @@ def pytorch_attention_enabled():
     global ENABLE_PYTORCH_ATTENTION
     return ENABLE_PYTORCH_ATTENTION
 
+def pytorch_attention_enabled_vae():
+    if is_amd():
+        return False  # enabling pytorch attention on AMD currently causes crash when doing high res
+    return pytorch_attention_enabled()
+
 def pytorch_attention_flash_attention():
     global ENABLE_PYTORCH_ATTENTION
     if ENABLE_PYTORCH_ATTENTION:
@@ -908,6 +936,8 @@ def pytorch_attention_flash_attention():
             return True
         if is_ascend_npu():
             return True
+        if is_amd():
+            return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention
     return False
 
 def mac_version():
@@ -920,11 +950,11 @@ def force_upcast_attention_dtype():
     upcast = args.force_upcast_attention
 
     macos_version = mac_version()
-    if macos_version is not None and ((14, 5) <= macos_version <= (15, 2)):  # black image bug on recent versions of macOS
+    if macos_version is not None and ((14, 5) <= macos_version < (16,)):  # black image bug on recent versions of macOS
         upcast = True
 
     if upcast:
-        return torch.float32
+        return {torch.float16: torch.float32}
     else:
         return None
 
@@ -990,21 +1020,26 @@ def is_device_mps(device):
 def is_device_cuda(device):
     return is_device_type(device, 'cuda')
 
-def should_use_fp16(device=None, model_params=0, prioritize_performance=True, manual_cast=False):
+def is_directml_enabled():
     global directml_enabled
+    if directml_enabled:
+        return True
 
+    return False
+
+def should_use_fp16(device=None, model_params=0, prioritize_performance=True, manual_cast=False):
     if device is not None:
         if is_device_cpu(device):
             return False
 
-    if FORCE_FP16:
+    if args.force_fp16:
         return True
 
     if FORCE_FP32:
         return False
 
-    if directml_enabled:
-        return False
+    if is_directml_enabled():
+        return True
 
     if (device is not None and is_device_mps(device)) or mps_mode():
         return True
@@ -1075,13 +1110,23 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     if is_intel_xpu():
         return True
 
+    if is_ascend_npu():
+        return True
+
+    if is_amd():
+        arch = torch.cuda.get_device_properties(device).gcnArchName
+        if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]):  # RDNA2 and older don't support bf16
+            if manual_cast:
+                return True
+            return False
+
     props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
 
     bf16_works = torch.cuda.is_bf16_supported()
 
-    if bf16_works or manual_cast:
+    if bf16_works and manual_cast:
         free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
@@ -1100,11 +1145,11 @@ def supports_fp8_compute(device=None):
     if props.minor < 9:
         return False
 
-    if int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) < 3):
+    if torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 3):
         return False
 
     if WINDOWS:
-        if (int(torch_version[0]) == 2 and int(torch_version[2]) < 4):
+        if (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 4):
             return False
 
     return True

comfy/model_patcher.py

@@ -96,8 +96,28 @@ def wipe_lowvram_weight(m):
     if hasattr(m, "prev_comfy_cast_weights"):
         m.comfy_cast_weights = m.prev_comfy_cast_weights
         del m.prev_comfy_cast_weights
-    m.weight_function = None
-    m.bias_function = None
+
+    if hasattr(m, "weight_function"):
+        m.weight_function = []
+
+    if hasattr(m, "bias_function"):
+        m.bias_function = []
+
+def move_weight_functions(m, device):
+    if device is None:
+        return 0
+
+    memory = 0
+    if hasattr(m, "weight_function"):
+        for f in m.weight_function:
+            if hasattr(f, "move_to"):
+                memory += f.move_to(device=device)
+
+    if hasattr(m, "bias_function"):
+        for f in m.bias_function:
+            if hasattr(f, "move_to"):
+                memory += f.move_to(device=device)
+    return memory
 
 class LowVramPatch:
     def __init__(self, key, patches):
@@ -192,11 +212,13 @@ class ModelPatcher:
         self.backup = {}
         self.object_patches = {}
         self.object_patches_backup = {}
+        self.weight_wrapper_patches = {}
         self.model_options = {"transformer_options":{}}
         self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
         self.weight_inplace_update = weight_inplace_update
+        self.force_cast_weights = False
         self.patches_uuid = uuid.uuid4()
         self.parent = None
 
@@ -250,11 +272,14 @@ class ModelPatcher:
         n.patches_uuid = self.patches_uuid
 
         n.object_patches = self.object_patches.copy()
+        n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
         n.model_options = copy.deepcopy(self.model_options)
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
 
+        n.force_cast_weights = self.force_cast_weights
+
         # attachments
         n.attachments = {}
         for k in self.attachments:
@@ -402,6 +427,16 @@ class ModelPatcher:
     def add_object_patch(self, name, obj):
         self.object_patches[name] = obj
 
+    def set_model_compute_dtype(self, dtype):
+        self.add_object_patch("manual_cast_dtype", dtype)
+        if dtype is not None:
+            self.force_cast_weights = True
+        self.patches_uuid = uuid.uuid4() #TODO: optimize by preventing a full model reload for this
+
+    def add_weight_wrapper(self, name, function):
+        self.weight_wrapper_patches[name] = self.weight_wrapper_patches.get(name, []) + [function]
+        self.patches_uuid = uuid.uuid4()
+
     def get_model_object(self, name: str) -> torch.nn.Module:
         """Retrieves a nested attribute from an object using dot notation considering
         object patches.
@@ -566,6 +601,9 @@ class ModelPatcher:
 
                 lowvram_weight = False
 
+                weight_key = "{}.weight".format(n)
+                bias_key = "{}.bias".format(n)
+
                 if not full_load and hasattr(m, "comfy_cast_weights"):
                     if mem_counter + module_mem >= lowvram_model_memory:
                         lowvram_weight = True
@@ -573,34 +611,46 @@ class ModelPatcher:
                         if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed
                             continue
 
-                weight_key = "{}.weight".format(n)
-                bias_key = "{}.bias".format(n)
-
+                cast_weight = self.force_cast_weights
                 if lowvram_weight:
+                    if hasattr(m, "comfy_cast_weights"):
+                        m.weight_function = []
+                        m.bias_function = []
+
                     if weight_key in self.patches:
                         if force_patch_weights:
                             self.patch_weight_to_device(weight_key)
                         else:
-                            m.weight_function = LowVramPatch(weight_key, self.patches)
+                            m.weight_function = [LowVramPatch(weight_key, self.patches)]
                             patch_counter += 1
                     if bias_key in self.patches:
                         if force_patch_weights:
                             self.patch_weight_to_device(bias_key)
                         else:
-                            m.bias_function = LowVramPatch(bias_key, self.patches)
+                            m.bias_function = [LowVramPatch(bias_key, self.patches)]
                             patch_counter += 1
 
-                    m.prev_comfy_cast_weights = m.comfy_cast_weights
-                    m.comfy_cast_weights = True
+                    cast_weight = True
                 else:
                     if hasattr(m, "comfy_cast_weights"):
-                        if m.comfy_cast_weights:
-                            wipe_lowvram_weight(m)
+                        wipe_lowvram_weight(m)
 
                     if full_load or mem_counter + module_mem < lowvram_model_memory:
                         mem_counter += module_mem
                         load_completely.append((module_mem, n, m, params))
 
+                if cast_weight and hasattr(m, "comfy_cast_weights"):
+                    m.prev_comfy_cast_weights = m.comfy_cast_weights
+                    m.comfy_cast_weights = True
+
+                if weight_key in self.weight_wrapper_patches:
+                    m.weight_function.extend(self.weight_wrapper_patches[weight_key])
+
+                if bias_key in self.weight_wrapper_patches:
+                    m.bias_function.extend(self.weight_wrapper_patches[bias_key])
+
+                mem_counter += move_weight_functions(m, device_to)
+
             load_completely.sort(reverse=True)
             for x in load_completely:
                 n = x[1]
@@ -662,6 +712,7 @@ class ModelPatcher:
             self.unpatch_hooks()
             if self.model.model_lowvram:
                 for m in self.model.modules():
+                    move_weight_functions(m, device_to)
                     wipe_lowvram_weight(m)
 
                 self.model.model_lowvram = False
@@ -728,15 +779,19 @@ class ModelPatcher:
                     weight_key = "{}.weight".format(n)
                     bias_key = "{}.bias".format(n)
                     if move_weight:
+                        cast_weight = self.force_cast_weights
                         m.to(device_to)
+                        module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                m.weight_function = LowVramPatch(weight_key, self.patches)
+                                m.weight_function.append(LowVramPatch(weight_key, self.patches))
                                 patch_counter += 1
                             if bias_key in self.patches:
-                                m.bias_function = LowVramPatch(bias_key, self.patches)
+                                m.bias_function.append(LowVramPatch(bias_key, self.patches))
                                 patch_counter += 1
+                            cast_weight = True
 
+                        if cast_weight:
                             m.prev_comfy_cast_weights = m.comfy_cast_weights
                             m.comfy_cast_weights = True
                         m.comfy_patched_weights = False

comfy/ops.py

@@ -38,21 +38,23 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
     bias = None
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
     if s.bias is not None:
-        has_function = s.bias_function is not None
+        has_function = len(s.bias_function) > 0
         bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
         if has_function:
-            bias = s.bias_function(bias)
+            for f in s.bias_function:
+                bias = f(bias)
 
-    has_function = s.weight_function is not None
+    has_function = len(s.weight_function) > 0
     weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
     if has_function:
-        weight = s.weight_function(weight)
+        for f in s.weight_function:
+            weight = f(weight)
     return weight, bias
 
 class CastWeightBiasOp:
     comfy_cast_weights = False
-    weight_function = None
-    bias_function = None
+    weight_function = []
+    bias_function = []
 
 class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
@@ -64,7 +66,7 @@ class disable_weight_init:
             return torch.nn.functional.linear(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -78,7 +80,7 @@ class disable_weight_init:
             return self._conv_forward(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -92,7 +94,7 @@ class disable_weight_init:
             return self._conv_forward(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -106,7 +108,7 @@ class disable_weight_init:
             return self._conv_forward(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -120,12 +122,11 @@ class disable_weight_init:
             return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
 
-
     class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp):
         def reset_parameters(self):
             return None
@@ -139,7 +140,7 @@ class disable_weight_init:
             return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -160,7 +161,7 @@ class disable_weight_init:
                 output_padding, self.groups, self.dilation)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -181,7 +182,7 @@ class disable_weight_init:
                 output_padding, self.groups, self.dilation)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -199,7 +200,7 @@ class disable_weight_init:
             return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 if "out_dtype" in kwargs:

comfy/samplers.py

@@ -686,7 +686,8 @@ class Sampler:
 KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
-                  "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "gradient_estimation"]
+                  "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
+                  "gradient_estimation"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):

comfy/sd.py

@@ -12,6 +12,7 @@ from .ldm.audio.autoencoder import AudioOobleckVAE
 import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
 import comfy.ldm.cosmos.vae
+import comfy.ldm.wan.vae
 import yaml
 import math
 
@@ -37,6 +38,7 @@ import comfy.text_encoders.lt
 import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
+import comfy.text_encoders.wan
 
 import comfy.model_patcher
 import comfy.lora
@@ -392,6 +394,18 @@ class VAE:
                 self.memory_used_decode = lambda shape, dtype: (50 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
                 self.memory_used_encode = lambda shape, dtype: (50 * (round((shape[2] + 7) / 8) * 8) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.working_dtypes = [torch.bfloat16, torch.float32]
+            elif "decoder.middle.0.residual.0.gamma" in sd:
+                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                self.upscale_index_formula = (4, 8, 8)
+                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                self.downscale_index_formula = (4, 8, 8)
+                self.latent_dim = 3
+                self.latent_channels = 16
+                ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
+                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -659,6 +673,7 @@ class CLIPType(Enum):
     PIXART = 10
     COSMOS = 11
     LUMINA2 = 12
+    WAN = 13
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -763,6 +778,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.PIXART:
                 clip_target.clip = comfy.text_encoders.pixart_t5.pixart_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.pixart_t5.PixArtTokenizer
+            elif clip_type == CLIPType.WAN:
+                clip_target.clip = comfy.text_encoders.wan.te(**t5xxl_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.wan.WanT5Tokenizer
+                tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer

comfy/supported_models.py

@@ -16,6 +16,7 @@ import comfy.text_encoders.lt
 import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
+import comfy.text_encoders.wan
 
 from . import supported_models_base
 from . import latent_formats
@@ -895,6 +896,49 @@ class Lumina2(supported_models_base.BASE):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}gemma2_2b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.lumina2.LuminaTokenizer, comfy.text_encoders.lumina2.te(**hunyuan_detect))
 
-models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2]
+class WAN21_T2V(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "t2v",
+    }
+
+    sampling_settings = {
+        "shift": 8.0,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Wan21
+
+    memory_usage_factor = 1.0
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = self.unet_config.get("dim", 2000) / 2000
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}umt5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.wan.WanT5Tokenizer, comfy.text_encoders.wan.te(**t5_detect))
+
+class WAN21_I2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "i2v",
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, image_to_video=True, device=device)
+        return out
+
+models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V]
 
 models += [SVD_img2vid]

comfy/text_encoders/lumina2.py

@@ -19,11 +19,6 @@ class LuminaTokenizer(sd1_clip.SD1Tokenizer):
 
 class Gemma2_2BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
-        llama_scaled_fp8 = model_options.get("llama_scaled_fp8", None)
-        if llama_scaled_fp8 is not None:
-            model_options = model_options.copy()
-            model_options["scaled_fp8"] = llama_scaled_fp8
-
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma2_2B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 
@@ -35,10 +30,10 @@ class LuminaModel(sd1_clip.SD1ClipModel):
 def te(dtype_llama=None, llama_scaled_fp8=None):
     class LuminaTEModel_(LuminaModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["llama_scaled_fp8"] = llama_scaled_fp8
-                if dtype_llama is not None:
-                    dtype = dtype_llama
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return LuminaTEModel_

comfy/text_encoders/umt5_config_xxl.json

@@ -0,0 +1,22 @@
+{
+  "d_ff": 10240,
+  "d_kv": 64,
+  "d_model": 4096,
+  "decoder_start_token_id": 0,
+  "dropout_rate": 0.1,
+  "eos_token_id": 1,
+  "dense_act_fn": "gelu_pytorch_tanh",
+  "initializer_factor": 1.0,
+  "is_encoder_decoder": true,
+  "is_gated_act": true,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "umt5",
+  "num_decoder_layers": 24,
+  "num_heads": 64,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "vocab_size": 256384
+}

comfy/text_encoders/wan.py

@@ -0,0 +1,37 @@
+from comfy import sd1_clip
+from .spiece_tokenizer import SPieceTokenizer
+import comfy.text_encoders.t5
+import os
+
+class UMT5XXlModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "umt5_config_xxl.json")
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True, model_options=model_options)
+
+class UMT5XXlTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=4096, embedding_key='umt5xxl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=0)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+
+class WanT5Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="umt5xxl", tokenizer=UMT5XXlTokenizer)
+
+class WanT5Model(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="umt5xxl", clip_model=UMT5XXlModel, **kwargs)
+
+def te(dtype_t5=None, t5xxl_scaled_fp8=None):
+    class WanTEModel(WanT5Model):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5xxl_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype_t5 is not None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return WanTEModel

comfy/utils.py

@@ -58,7 +58,7 @@ def load_torch_file(ckpt, safe_load=False, device=None):
                 if "HeaderTooLarge" in message:
                     raise ValueError("{}\n\nFile path: {}\n\nThe safetensors file is corrupt or invalid. Make sure this is actually a safetensors file and not a ckpt or pt or other filetype.".format(message, ckpt))
                 if "MetadataIncompleteBuffer" in message:
-                    raise ValueError("{}\n\nFile path: {}\n\nThe safetensors file is incomplete. Check the file size and make sure you have copied/downloaded it correctly.".format(message, ckpt))
+                    raise ValueError("{}\n\nFile path: {}\n\nThe safetensors file is corrupt/incomplete. Check the file size and make sure you have copied/downloaded it correctly.".format(message, ckpt))
             raise e
     else:
         if safe_load or ALWAYS_SAFE_LOAD:

comfy_extras/nodes_load_3d.py

@@ -20,9 +20,7 @@ class Load3D():
             "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
             "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
             "material": (["original", "normal", "wireframe", "depth"],),
-            "light_intensity": ("INT", {"default": 10, "min": 1, "max": 20, "step": 1}),
             "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
-            "fov": ("INT", {"default": 75, "min": 10, "max": 150, "step": 1}),
         }}
 
     RETURN_TYPES = ("IMAGE", "MASK", "STRING")
@@ -34,22 +32,14 @@ class Load3D():
     CATEGORY = "3d"
 
     def process(self, model_file, image, **kwargs):
-        if isinstance(image, dict):
-            image_path = folder_paths.get_annotated_filepath(image['image'])
-            mask_path = folder_paths.get_annotated_filepath(image['mask'])
+        image_path = folder_paths.get_annotated_filepath(image['image'])
+        mask_path = folder_paths.get_annotated_filepath(image['mask'])
 
-            load_image_node = nodes.LoadImage()
-            output_image, ignore_mask = load_image_node.load_image(image=image_path)
-            ignore_image, output_mask = load_image_node.load_image(image=mask_path)
+        load_image_node = nodes.LoadImage()
+        output_image, ignore_mask = load_image_node.load_image(image=image_path)
+        ignore_image, output_mask = load_image_node.load_image(image=mask_path)
 
-            return output_image, output_mask, model_file,
-        else:
-            # to avoid the format is not dict which will happen the FE code is not compatibility to core,
-            # we need to this to double-check, it can be removed after merged FE into the core
-            image_path = folder_paths.get_annotated_filepath(image)
-            load_image_node = nodes.LoadImage()
-            output_image, output_mask = load_image_node.load_image(image=image_path)
-            return output_image, output_mask, model_file,
+        return output_image, output_mask, model_file,
 
 class Load3DAnimation():
     @classmethod
@@ -66,9 +56,7 @@ class Load3DAnimation():
             "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
             "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
             "material": (["original", "normal", "wireframe", "depth"],),
-            "light_intensity": ("INT", {"default": 10, "min": 1, "max": 20, "step": 1}),
             "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
-            "fov": ("INT", {"default": 75, "min": 10, "max": 150, "step": 1}),
         }}
 
     RETURN_TYPES = ("IMAGE", "MASK", "STRING")
@@ -80,20 +68,14 @@ class Load3DAnimation():
     CATEGORY = "3d"
 
     def process(self, model_file, image, **kwargs):
-        if isinstance(image, dict):
-            image_path = folder_paths.get_annotated_filepath(image['image'])
-            mask_path = folder_paths.get_annotated_filepath(image['mask'])
+        image_path = folder_paths.get_annotated_filepath(image['image'])
+        mask_path = folder_paths.get_annotated_filepath(image['mask'])
 
-            load_image_node = nodes.LoadImage()
-            output_image, ignore_mask = load_image_node.load_image(image=image_path)
-            ignore_image, output_mask = load_image_node.load_image(image=mask_path)
+        load_image_node = nodes.LoadImage()
+        output_image, ignore_mask = load_image_node.load_image(image=image_path)
+        ignore_image, output_mask = load_image_node.load_image(image=mask_path)
 
-            return output_image, output_mask, model_file,
-        else:
-            image_path = folder_paths.get_annotated_filepath(image)
-            load_image_node = nodes.LoadImage()
-            output_image, output_mask = load_image_node.load_image(image=image_path)
-            return output_image, output_mask, model_file,
+        return output_image, output_mask, model_file,
 
 class Preview3D():
     @classmethod
@@ -101,9 +83,7 @@ class Preview3D():
         return {"required": {
             "model_file": ("STRING", {"default": "", "multiline": False}),
             "material": (["original", "normal", "wireframe", "depth"],),
-            "light_intensity": ("INT", {"default": 10, "min": 1, "max": 20, "step": 1}),
             "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
-            "fov": ("INT", {"default": 75, "min": 10, "max": 150, "step": 1}),
         }}
 
     OUTPUT_NODE = True
@@ -123,9 +103,7 @@ class Preview3DAnimation():
         return {"required": {
             "model_file": ("STRING", {"default": "", "multiline": False}),
             "material": (["original", "normal", "wireframe", "depth"],),
-            "light_intensity": ("INT", {"default": 10, "min": 1, "max": 20, "step": 1}),
             "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
-            "fov": ("INT", {"default": 75, "min": 10, "max": 150, "step": 1}),
         }}
 
     OUTPUT_NODE = True

comfy_extras/nodes_lumina2.py

@@ -0,0 +1,104 @@
+from comfy.comfy_types import IO, ComfyNodeABC, InputTypeDict
+import torch
+
+
+class RenormCFG:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "cfg_trunc": ("FLOAT", {"default": 100, "min": 0.0, "max": 100.0, "step": 0.01}),
+                              "renorm_cfg": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01}),
+                              }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/model"
+
+    def patch(self, model, cfg_trunc, renorm_cfg):
+        def renorm_cfg_func(args):
+            cond_denoised = args["cond_denoised"]
+            uncond_denoised = args["uncond_denoised"]
+            cond_scale = args["cond_scale"]
+            timestep = args["timestep"]
+            x_orig = args["input"]
+            in_channels = model.model.diffusion_model.in_channels
+
+            if timestep[0] < cfg_trunc:
+                cond_eps, uncond_eps = cond_denoised[:, :in_channels], uncond_denoised[:, :in_channels]
+                cond_rest, _ = cond_denoised[:, in_channels:], uncond_denoised[:, in_channels:]
+                half_eps = uncond_eps + cond_scale * (cond_eps - uncond_eps)
+                half_rest = cond_rest
+
+                if float(renorm_cfg) > 0.0:
+                    ori_pos_norm = torch.linalg.vector_norm(cond_eps
+                            , dim=tuple(range(1, len(cond_eps.shape))), keepdim=True
+                    )
+                    max_new_norm = ori_pos_norm * float(renorm_cfg)
+                    new_pos_norm = torch.linalg.vector_norm(
+                            half_eps, dim=tuple(range(1, len(half_eps.shape))), keepdim=True
+                        )
+                    if new_pos_norm >= max_new_norm:
+                        half_eps = half_eps * (max_new_norm / new_pos_norm)
+            else:
+                cond_eps, uncond_eps = cond_denoised[:, :in_channels], uncond_denoised[:, :in_channels]
+                cond_rest, _ = cond_denoised[:, in_channels:], uncond_denoised[:, in_channels:]
+                half_eps = cond_eps
+                half_rest = cond_rest
+
+            cfg_result = torch.cat([half_eps, half_rest], dim=1)
+
+            # cfg_result = uncond_denoised + (cond_denoised - uncond_denoised) * cond_scale
+
+            return x_orig - cfg_result
+
+        m = model.clone()
+        m.set_model_sampler_cfg_function(renorm_cfg_func)
+        return (m, )
+
+
+class CLIPTextEncodeLumina2(ComfyNodeABC):
+    SYSTEM_PROMPT = {
+        "superior": "You are an assistant designed to generate superior images with the superior "\
+            "degree of image-text alignment based on textual prompts or user prompts.",
+        "alignment": "You are an assistant designed to generate high-quality images with the "\
+            "highest degree of image-text alignment based on textual prompts."
+    }
+    SYSTEM_PROMPT_TIP = "Lumina2 provide two types of system prompts:" \
+        "Superior: You are an assistant designed to generate superior images with the superior "\
+        "degree of image-text alignment based on textual prompts or user prompts. "\
+        "Alignment: You are an assistant designed to generate high-quality images with the highest "\
+        "degree of image-text alignment based on textual prompts."
+    @classmethod
+    def INPUT_TYPES(s) -> InputTypeDict:
+        return {
+            "required": {
+                "system_prompt": (list(CLIPTextEncodeLumina2.SYSTEM_PROMPT.keys()), {"tooltip": CLIPTextEncodeLumina2.SYSTEM_PROMPT_TIP}),
+                "user_prompt": (IO.STRING, {"multiline": True, "dynamicPrompts": True, "tooltip": "The text to be encoded."}),
+                "clip": (IO.CLIP, {"tooltip": "The CLIP model used for encoding the text."})
+            }
+        }
+    RETURN_TYPES = (IO.CONDITIONING,)
+    OUTPUT_TOOLTIPS = ("A conditioning containing the embedded text used to guide the diffusion model.",)
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning"
+    DESCRIPTION = "Encodes a system prompt and a user prompt using a CLIP model into an embedding that can be used to guide the diffusion model towards generating specific images."
+
+    def encode(self, clip, user_prompt, system_prompt):
+        if clip is None:
+            raise RuntimeError("ERROR: clip input is invalid: None\n\nIf the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.")
+        system_prompt = CLIPTextEncodeLumina2.SYSTEM_PROMPT[system_prompt]
+        prompt = f'{system_prompt} <Prompt Start> {user_prompt}'
+        tokens = clip.tokenize(prompt)
+        return (clip.encode_from_tokens_scheduled(tokens), )
+
+
+NODE_CLASS_MAPPINGS = {
+    "CLIPTextEncodeLumina2": CLIPTextEncodeLumina2,
+    "RenormCFG": RenormCFG
+}
+
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "CLIPTextEncodeLumina2": "CLIP Text Encode for Lumina2",
+}

comfy_extras/nodes_model_advanced.py

@@ -3,6 +3,8 @@ import comfy.model_sampling
 import comfy.latent_formats
 import nodes
 import torch
+import node_helpers
+
 
 class LCM(comfy.model_sampling.EPS):
     def calculate_denoised(self, sigma, model_output, model_input):
@@ -294,6 +296,24 @@ class RescaleCFG:
         m.set_model_sampler_cfg_function(rescale_cfg)
         return (m, )
 
+class ModelComputeDtype:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "dtype": (["default", "fp32", "fp16", "bf16"],),
+                              }}
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/debug/model"
+
+    def patch(self, model, dtype):
+        m = model.clone()
+        m.set_model_compute_dtype(node_helpers.string_to_torch_dtype(dtype))
+        return (m, )
+
+
 NODE_CLASS_MAPPINGS = {
     "ModelSamplingDiscrete": ModelSamplingDiscrete,
     "ModelSamplingContinuousEDM": ModelSamplingContinuousEDM,
@@ -303,4 +323,5 @@ NODE_CLASS_MAPPINGS = {
     "ModelSamplingAuraFlow": ModelSamplingAuraFlow,
     "ModelSamplingFlux": ModelSamplingFlux,
     "RescaleCFG": RescaleCFG,
+    "ModelComputeDtype": ModelComputeDtype,
 }

comfy_extras/nodes_video.py

@@ -0,0 +1,76 @@
+import os
+import av
+import torch
+import folder_paths
+import json
+from fractions import Fraction
+
+
+class SaveWEBM:
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+        self.type = "output"
+        self.prefix_append = ""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"images": ("IMAGE", ),
+                     "filename_prefix": ("STRING", {"default": "ComfyUI"}),
+                     "codec": (["vp9", "av1"],),
+                     "fps": ("FLOAT", {"default": 24.0, "min": 0.01, "max": 1000.0, "step": 0.01}),
+                     "crf": ("FLOAT", {"default": 32.0, "min": 0, "max": 63.0, "step": 1, "tooltip": "Higher crf means lower quality with a smaller file size, lower crf means higher quality higher filesize."}),
+                     },
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},
+                }
+
+    RETURN_TYPES = ()
+    FUNCTION = "save_images"
+
+    OUTPUT_NODE = True
+
+    CATEGORY = "image/video"
+
+    EXPERIMENTAL = True
+
+    def save_images(self, images, codec, fps, filename_prefix, crf, prompt=None, extra_pnginfo=None):
+        filename_prefix += self.prefix_append
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0])
+
+        file = f"{filename}_{counter:05}_.webm"
+        container = av.open(os.path.join(full_output_folder, file), mode="w")
+
+        if prompt is not None:
+            container.metadata["prompt"] = json.dumps(prompt)
+
+        if extra_pnginfo is not None:
+            for x in extra_pnginfo:
+                container.metadata[x] = json.dumps(extra_pnginfo[x])
+
+        codec_map = {"vp9": "libvpx-vp9", "av1": "libaom-av1"}
+        stream = container.add_stream(codec_map[codec], rate=Fraction(round(fps * 1000), 1000))
+        stream.width = images.shape[-2]
+        stream.height = images.shape[-3]
+        stream.pix_fmt = "yuv420p"
+        stream.bit_rate = 0
+        stream.options = {'crf': str(crf)}
+
+        for frame in images:
+            frame = av.VideoFrame.from_ndarray(torch.clamp(frame[..., :3] * 255, min=0, max=255).to(device=torch.device("cpu"), dtype=torch.uint8).numpy(), format="rgb24")
+            for packet in stream.encode(frame):
+                container.mux(packet)
+        container.mux(stream.encode())
+        container.close()
+
+        results = [{
+            "filename": file,
+            "subfolder": subfolder,
+            "type": self.type
+        }]
+
+        return {"ui": {"images": results, "animated": (True,)}}  # TODO: frontend side
+
+
+NODE_CLASS_MAPPINGS = {
+    "SaveWEBM": SaveWEBM,
+}

comfy_extras/nodes_wan.py

@@ -0,0 +1,54 @@
+import nodes
+import node_helpers
+import torch
+import comfy.model_management
+import comfy.utils
+
+
+class WanImageToVideo:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"positive": ("CONDITIONING", ),
+                             "negative": ("CONDITIONING", ),
+                             "vae": ("VAE", ),
+                             "width": ("INT", {"default": 1280, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "height": ("INT", {"default": 720, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "length": ("INT", {"default": 121, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                },
+                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
+                             "start_image": ("IMAGE", ),
+                }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
+    RETURN_NAMES = ("positive", "negative", "latent")
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None):
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            image = torch.ones((length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5
+            image[:start_image.shape[0]] = start_image
+
+            concat_latent_image = vae.encode(image[:, :, :, :3])
+            mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype)
+            mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0
+
+            positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return (positive, negative, out_latent)
+
+
+NODE_CLASS_MAPPINGS = {
+    "WanImageToVideo": WanImageToVideo,
+}

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.14"
+__version__ = "0.3.16"

node_helpers.py

@@ -1,4 +1,5 @@
 import hashlib
+import torch
 
 from comfy.cli_args import args
 
@@ -35,3 +36,11 @@ def hasher():
         "sha512": hashlib.sha512
     }
     return hashfuncs[args.default_hashing_function]
+
+def string_to_torch_dtype(string):
+    if string == "fp32":
+        return torch.float32
+    if string == "fp16":
+        return torch.float16
+    if string == "bf16":
+        return torch.bfloat16

nodes.py

@@ -914,7 +914,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),
@@ -924,7 +924,7 @@ class CLIPLoader:
 
     CATEGORY = "advanced/loaders"
 
-    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 / clip-g / clip-l\nstable_audio: t5\nmochi: t5\ncosmos: old t5 xxl"
+    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl"
 
     def load_clip(self, clip_name, type="stable_diffusion", device="default"):
         if type == "stable_cascade":
@@ -943,6 +943,8 @@ class CLIPLoader:
             clip_type = comfy.sd.CLIPType.COSMOS
         elif type == "lumina2":
             clip_type = comfy.sd.CLIPType.LUMINA2
+        elif type == "wan":
+            clip_type = comfy.sd.CLIPType.WAN
         else:
             clip_type = comfy.sd.CLIPType.STABLE_DIFFUSION
 
@@ -1064,10 +1066,11 @@ class StyleModelApply:
         for t in conditioning:
             (txt, keys) = t
             keys = keys.copy()
-            if strength_type == "attn_bias" and strength != 1.0:
+            # even if the strength is 1.0 (i.e, no change), if there's already a mask, we have to add to it
+            if "attention_mask" in keys or (strength_type == "attn_bias" and strength != 1.0):
                 # math.log raises an error if the argument is zero
                 # torch.log returns -inf, which is what we want
-                attn_bias = torch.log(torch.Tensor([strength]))
+                attn_bias = torch.log(torch.Tensor([strength if strength_type == "attn_bias" else 1.0]))
                 # get the size of the mask image
                 mask_ref_size = keys.get("attention_mask_img_shape", (1, 1))
                 n_ref = mask_ref_size[0] * mask_ref_size[1]
@@ -1762,6 +1765,36 @@ class LoadImageMask:
 
         return True
 
+
+class LoadImageOutput(LoadImage):
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image": ("COMBO", {
+                    "image_upload": True,
+                    "image_folder": "output",
+                    "remote": {
+                        "route": "/internal/files/output",
+                        "refresh_button": True,
+                        "control_after_refresh": "first",
+                    },
+                }),
+            }
+        }
+
+    DESCRIPTION = "Load an image from the output folder. When the refresh button is clicked, the node will update the image list and automatically select the first image, allowing for easy iteration."
+    EXPERIMENTAL = True
+    FUNCTION = "load_image_output"
+
+    def load_image_output(self, image):
+        return self.load_image(f"{image} [output]")
+
+    @classmethod
+    def VALIDATE_INPUTS(s, image):
+        return True
+
+
 class ImageScale:
     upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"]
     crop_methods = ["disabled", "center"]
@@ -1948,6 +1981,7 @@ NODE_CLASS_MAPPINGS = {
     "PreviewImage": PreviewImage,
     "LoadImage": LoadImage,
     "LoadImageMask": LoadImageMask,
+    "LoadImageOutput": LoadImageOutput,
     "ImageScale": ImageScale,
     "ImageScaleBy": ImageScaleBy,
     "ImageInvert": ImageInvert,
@@ -2048,6 +2082,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "PreviewImage": "Preview Image",
     "LoadImage": "Load Image",
     "LoadImageMask": "Load Image (as Mask)",
+    "LoadImageOutput": "Load Image (from Outputs)",
     "ImageScale": "Upscale Image",
     "ImageScaleBy": "Upscale Image By",
     "ImageUpscaleWithModel": "Upscale Image (using Model)",
@@ -2232,6 +2267,9 @@ def init_builtin_extra_nodes():
         "nodes_hooks.py",
         "nodes_load_3d.py",
         "nodes_cosmos.py",
+        "nodes_video.py",
+        "nodes_lumina2.py",
+        "nodes_wan.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.14"
+version = "0.3.16"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -8,7 +8,8 @@ transformers>=4.28.1
 tokenizers>=0.13.3
 sentencepiece
 safetensors>=0.4.2
-aiohttp
+aiohttp>=3.11.8
+yarl>=1.18.0
 pyyaml
 Pillow
 scipy
@@ -19,3 +20,4 @@ psutil
 kornia>=0.7.1
 spandrel
 soundfile
+av

server.py

@@ -57,8 +57,6 @@ async def cache_control(request: web.Request, handler):
 async def compress_body(request: web.Request, handler):
     accept_encoding = request.headers.get("Accept-Encoding", "")
     response: web.Response = await handler(request)
-    if args.disable_compres_response_body:
-        return response
     if not isinstance(response, web.Response):
         return response
     if response.content_type not in ["application/json", "text/plain"]:
@@ -152,7 +150,8 @@ class PromptServer():
         PromptServer.instance = self
 
         mimetypes.init()
-        mimetypes.types_map['.js'] = 'application/javascript; charset=utf-8'
+        mimetypes.add_type('application/javascript; charset=utf-8', '.js')
+        mimetypes.add_type('image/webp', '.webp')
 
         self.user_manager = UserManager()
         self.model_file_manager = ModelFileManager()
@@ -165,7 +164,10 @@ class PromptServer():
         self.client_session:Optional[aiohttp.ClientSession] = None
         self.number = 0
 
-        middlewares = [cache_control, compress_body]
+        middlewares = [cache_control]
+        if args.enable_compress_response_body:
+            middlewares.append(compress_body)
+
         if args.enable_cors_header:
             middlewares.append(create_cors_middleware(args.enable_cors_header))
         else:

tests-unit/server/routes/internal_routes_test.py

@@ -1,115 +0,0 @@
-import pytest
-from aiohttp import web
-from unittest.mock import MagicMock, patch
-from api_server.routes.internal.internal_routes import InternalRoutes
-from api_server.services.file_service import FileService
-from folder_paths import models_dir, user_directory, output_directory
-
-
-@pytest.fixture
-def internal_routes():
-    return InternalRoutes(None)
-
-@pytest.fixture
-def aiohttp_client_factory(aiohttp_client, internal_routes):
-    async def _get_client():
-        app = internal_routes.get_app()
-        return await aiohttp_client(app)
-    return _get_client
-
-@pytest.mark.asyncio
-async def test_list_files_valid_directory(aiohttp_client_factory, internal_routes):
-    mock_file_list = [
-        {"name": "file1.txt", "path": "file1.txt", "type": "file", "size": 100},
-        {"name": "dir1", "path": "dir1", "type": "directory"}
-    ]
-    internal_routes.file_service.list_files = MagicMock(return_value=mock_file_list)
-    client = await aiohttp_client_factory()
-    resp = await client.get('/files?directory=models')
-    assert resp.status == 200
-    data = await resp.json()
-    assert 'files' in data
-    assert len(data['files']) == 2
-    assert data['files'] == mock_file_list
-
-    # Check other valid directories
-    resp = await client.get('/files?directory=user')
-    assert resp.status == 200
-    resp = await client.get('/files?directory=output')
-    assert resp.status == 200
-
-@pytest.mark.asyncio
-async def test_list_files_invalid_directory(aiohttp_client_factory, internal_routes):
-    internal_routes.file_service.list_files = MagicMock(side_effect=ValueError("Invalid directory key"))
-    client = await aiohttp_client_factory()
-    resp = await client.get('/files?directory=invalid')
-    assert resp.status == 400
-    data = await resp.json()
-    assert 'error' in data
-    assert data['error'] == "Invalid directory key"
-
-@pytest.mark.asyncio
-async def test_list_files_exception(aiohttp_client_factory, internal_routes):
-    internal_routes.file_service.list_files = MagicMock(side_effect=Exception("Unexpected error"))
-    client = await aiohttp_client_factory()
-    resp = await client.get('/files?directory=models')
-    assert resp.status == 500
-    data = await resp.json()
-    assert 'error' in data
-    assert data['error'] == "Unexpected error"
-
-@pytest.mark.asyncio
-async def test_list_files_no_directory_param(aiohttp_client_factory, internal_routes):
-    mock_file_list = []
-    internal_routes.file_service.list_files = MagicMock(return_value=mock_file_list)
-    client = await aiohttp_client_factory()
-    resp = await client.get('/files')
-    assert resp.status == 200
-    data = await resp.json()
-    assert 'files' in data
-    assert len(data['files']) == 0
-
-def test_setup_routes(internal_routes):
-    internal_routes.setup_routes()
-    routes = internal_routes.routes
-    assert any(route.method == 'GET' and str(route.path) == '/files' for route in routes)
-
-def test_get_app(internal_routes):
-    app = internal_routes.get_app()
-    assert isinstance(app, web.Application)
-    assert internal_routes._app is not None
-
-def test_get_app_reuse(internal_routes):
-    app1 = internal_routes.get_app()
-    app2 = internal_routes.get_app()
-    assert app1 is app2
-
-@pytest.mark.asyncio
-async def test_routes_added_to_app(aiohttp_client_factory, internal_routes):
-    client = await aiohttp_client_factory()
-    try:
-        resp = await client.get('/files')
-        print(f"Response received: status {resp.status}")  # noqa: T201
-    except Exception as e:
-        print(f"Exception occurred during GET request: {e}")  # noqa: T201
-        raise
-
-    assert resp.status != 404, "Route /files does not exist"
-
-@pytest.mark.asyncio
-async def test_file_service_initialization():
-    with patch('api_server.routes.internal.internal_routes.FileService') as MockFileService:
-        # Create a mock instance
-        mock_file_service_instance = MagicMock(spec=FileService)
-        MockFileService.return_value = mock_file_service_instance
-        internal_routes = InternalRoutes(None)
-
-        # Check if FileService was initialized with the correct parameters
-        MockFileService.assert_called_once_with({
-            "models": models_dir,
-            "user": user_directory,
-            "output": output_directory
-        })
-
-        # Verify that the file_service attribute of InternalRoutes is set
-        assert internal_routes.file_service == mock_file_service_instance

tests-unit/server/services/file_service_test.py

@@ -1,54 +0,0 @@
-import pytest
-from unittest.mock import MagicMock
-from api_server.services.file_service import FileService
-
-@pytest.fixture
-def mock_file_system_ops():
-    return MagicMock()
-
-@pytest.fixture
-def file_service(mock_file_system_ops):
-    allowed_directories = {
-        "models": "/path/to/models",
-        "user": "/path/to/user",
-        "output": "/path/to/output"
-    }
-    return FileService(allowed_directories, file_system_ops=mock_file_system_ops)
-
-def test_list_files_valid_directory(file_service, mock_file_system_ops):
-    mock_file_system_ops.walk_directory.return_value = [
-        {"name": "file1.txt", "path": "file1.txt", "type": "file", "size": 100},
-        {"name": "dir1", "path": "dir1", "type": "directory"}
-    ]
-
-    result = file_service.list_files("models")
-
-    assert len(result) == 2
-    assert result[0]["name"] == "file1.txt"
-    assert result[1]["name"] == "dir1"
-    mock_file_system_ops.walk_directory.assert_called_once_with("/path/to/models")
-
-def test_list_files_invalid_directory(file_service):
-    # Does not support walking directories outside of the allowed directories
-    with pytest.raises(ValueError, match="Invalid directory key"):
-        file_service.list_files("invalid_key")
-
-def test_list_files_empty_directory(file_service, mock_file_system_ops):
-    mock_file_system_ops.walk_directory.return_value = []
-
-    result = file_service.list_files("models")
-
-    assert len(result) == 0
-    mock_file_system_ops.walk_directory.assert_called_once_with("/path/to/models")
-
-@pytest.mark.parametrize("directory_key", ["models", "user", "output"])
-def test_list_files_all_allowed_directories(file_service, mock_file_system_ops, directory_key):
-    mock_file_system_ops.walk_directory.return_value = [
-        {"name": f"file_{directory_key}.txt", "path": f"file_{directory_key}.txt", "type": "file", "size": 100}
-    ]
-
-    result = file_service.list_files(directory_key)
-
-    assert len(result) == 1
-    assert result[0]["name"] == f"file_{directory_key}.txt"
-    mock_file_system_ops.walk_directory.assert_called_once_with(f"/path/to/{directory_key}")

tests-unit/utils/extra_config_test.py

@@ -114,7 +114,7 @@ def test_load_extra_model_paths_expands_userpath(
     mock_yaml_safe_load.assert_called_once()
 
     # Check if open was called with the correct file path
-    mock_file.assert_called_once_with(dummy_yaml_file_name, 'r')
+    mock_file.assert_called_once_with(dummy_yaml_file_name, 'r', encoding='utf-8')
 
 
 @patch('builtins.open', new_callable=mock_open)
@@ -145,7 +145,7 @@ def test_load_extra_model_paths_expands_appdata(
     else:
         expected_base_path = '/Users/TestUser/AppData/Roaming/ComfyUI'
     expected_calls = [
-        ('checkpoints', os.path.join(expected_base_path, 'models/checkpoints'), False),
+        ('checkpoints', os.path.normpath(os.path.join(expected_base_path, 'models/checkpoints')), False),
     ]
 
     assert mock_add_model_folder_path.call_count == len(expected_calls)
@@ -197,8 +197,8 @@ def test_load_extra_path_config_relative_base_path(
 
     load_extra_path_config(dummy_yaml_name)
 
-    expected_checkpoints = os.path.abspath(os.path.join(str(tmp_path), sub_folder, "checkpoints"))
-    expected_some_value = os.path.abspath(os.path.join(str(tmp_path), sub_folder, "some_value"))
+    expected_checkpoints = os.path.abspath(os.path.join(str(tmp_path), "my_rel_base", "checkpoints"))
+    expected_some_value = os.path.abspath(os.path.join(str(tmp_path), "my_rel_base", "some_value"))
 
     actual_paths = folder_paths.folder_names_and_paths["checkpoints"][0]
     assert len(actual_paths) == 1, "Should have one path added for 'checkpoints'."

utils/extra_config.py

@@ -4,7 +4,7 @@ import folder_paths
 import logging
 
 def load_extra_path_config(yaml_path):
-    with open(yaml_path, 'r') as stream:
+    with open(yaml_path, 'r', encoding='utf-8') as stream:
         config = yaml.safe_load(stream)
     yaml_dir = os.path.dirname(os.path.abspath(yaml_path))
     for c in config:
@@ -29,5 +29,6 @@ def load_extra_path_config(yaml_path):
                     full_path = os.path.join(base_path, full_path)
                 elif not os.path.isabs(full_path):
                     full_path = os.path.abspath(os.path.join(yaml_dir, y))
-                logging.info("Adding extra search path {} {}".format(x, full_path))
-                folder_paths.add_model_folder_path(x, full_path, is_default)
+                normalized_path = os.path.normpath(full_path)
+                logging.info("Adding extra search path {} {}".format(x, normalized_path))
+                folder_paths.add_model_folder_path(x, normalized_path, is_default)

web/assets/BaseViewTemplate-BTbuZf5t.js

@@ -1,4 +1,4 @@
-import { d as defineComponent, U as ref, p as onMounted, b4 as isElectron, W as nextTick, b5 as electronAPI, o as openBlock, f as createElementBlock, i as withDirectives, v as vShow, j as unref, b6 as isNativeWindow, m as createBaseVNode, A as renderSlot, ai as normalizeClass } from "./index-4Hb32CNk.js";
+import { d as defineComponent, T as ref, p as onMounted, b8 as isElectron, V as nextTick, b9 as electronAPI, o as openBlock, f as createElementBlock, i as withDirectives, v as vShow, j as unref, ba as isNativeWindow, m as createBaseVNode, A as renderSlot, aj as normalizeClass } from "./index-Bv0b06LE.js";
 const _hoisted_1 = { class: "flex-grow w-full flex items-center justify-center overflow-auto" };
 const _sfc_main = /* @__PURE__ */ defineComponent({
   __name: "BaseViewTemplate",
@@ -27,7 +27,7 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
     });
     return (_ctx, _cache) => {
       return openBlock(), createElementBlock("div", {
-        class: normalizeClass(["font-sans w-screen h-screen flex flex-col pointer-events-auto", [
+        class: normalizeClass(["font-sans w-screen h-screen flex flex-col", [
           props.dark ? "text-neutral-300 bg-neutral-900 dark-theme" : "text-neutral-900 bg-neutral-300"
         ]])
       }, [
@@ -48,4 +48,4 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
 export {
   _sfc_main as _
 };
-//# sourceMappingURL=BaseViewTemplate-v6omkdXg.js.map
+//# sourceMappingURL=BaseViewTemplate-BTbuZf5t.js.map

web/assets/DesktopStartView-D9r53Bue.js

@@ -0,0 +1,19 @@
+import { d as defineComponent, o as openBlock, y as createBlock, z as withCtx, k as createVNode, j as unref, bE as script } from "./index-Bv0b06LE.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
+const _sfc_main = /* @__PURE__ */ defineComponent({
+  __name: "DesktopStartView",
+  setup(__props) {
+    return (_ctx, _cache) => {
+      return openBlock(), createBlock(_sfc_main$1, { dark: "" }, {
+        default: withCtx(() => [
+          createVNode(unref(script), { class: "m-8 w-48 h-48" })
+        ]),
+        _: 1
+      });
+    };
+  }
+});
+export {
+  _sfc_main as default
+};
+//# sourceMappingURL=DesktopStartView-D9r53Bue.js.map

web/assets/DesktopStartView-coDnSXEF.js

@@ -1,22 +0,0 @@
-import { d as defineComponent, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, k as createVNode, j as unref, bz as script } from "./index-4Hb32CNk.js";
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
-const _hoisted_1 = { class: "max-w-screen-sm w-screen p-8" };
-const _sfc_main = /* @__PURE__ */ defineComponent({
-  __name: "DesktopStartView",
-  setup(__props) {
-    return (_ctx, _cache) => {
-      return openBlock(), createBlock(_sfc_main$1, { dark: "" }, {
-        default: withCtx(() => [
-          createBaseVNode("div", _hoisted_1, [
-            createVNode(unref(script), { mode: "indeterminate" })
-          ])
-        ]),
-        _: 1
-      });
-    };
-  }
-});
-export {
-  _sfc_main as default
-};
-//# sourceMappingURL=DesktopStartView-coDnSXEF.js.map

web/assets/DesktopUpdateView-C-R0415K.js

@@ -0,0 +1,58 @@
+var __defProp = Object.defineProperty;
+var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
+import { d as defineComponent, T as ref, d8 as onUnmounted, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, j as unref, bg as t, k as createVNode, bE as script, l as script$1, b9 as electronAPI, _ as _export_sfc } from "./index-Bv0b06LE.js";
+import { s as script$2 } from "./index-A_bXPJCN.js";
+import { _ as _sfc_main$1 } from "./TerminalOutputDrawer-CKr7Br7O.js";
+import { _ as _sfc_main$2 } from "./BaseViewTemplate-BTbuZf5t.js";
+const _hoisted_1 = { class: "h-screen w-screen grid items-center justify-around overflow-y-auto" };
+const _hoisted_2 = { class: "relative m-8 text-center" };
+const _hoisted_3 = { class: "download-bg pi-download text-4xl font-bold" };
+const _hoisted_4 = { class: "m-8" };
+const _sfc_main = /* @__PURE__ */ defineComponent({
+  __name: "DesktopUpdateView",
+  setup(__props) {
+    const electron = electronAPI();
+    const terminalVisible = ref(false);
+    const toggleConsoleDrawer = /* @__PURE__ */ __name(() => {
+      terminalVisible.value = !terminalVisible.value;
+    }, "toggleConsoleDrawer");
+    onUnmounted(() => electron.Validation.dispose());
+    return (_ctx, _cache) => {
+      return openBlock(), createBlock(_sfc_main$2, { dark: "" }, {
+        default: withCtx(() => [
+          createBaseVNode("div", _hoisted_1, [
+            createBaseVNode("div", _hoisted_2, [
+              createBaseVNode("h1", _hoisted_3, toDisplayString(unref(t)("desktopUpdate.title")), 1),
+              createBaseVNode("div", _hoisted_4, [
+                createBaseVNode("span", null, toDisplayString(unref(t)("desktopUpdate.description")), 1)
+              ]),
+              createVNode(unref(script), { class: "m-8 w-48 h-48" }),
+              createVNode(unref(script$1), {
+                style: { "transform": "translateX(-50%)" },
+                class: "fixed bottom-0 left-1/2 my-8",
+                label: unref(t)("maintenance.consoleLogs"),
+                icon: "pi pi-desktop",
+                "icon-pos": "left",
+                severity: "secondary",
+                onClick: toggleConsoleDrawer
+              }, null, 8, ["label"]),
+              createVNode(_sfc_main$1, {
+                modelValue: terminalVisible.value,
+                "onUpdate:modelValue": _cache[0] || (_cache[0] = ($event) => terminalVisible.value = $event),
+                header: unref(t)("g.terminal"),
+                "default-message": unref(t)("desktopUpdate.terminalDefaultMessage")
+              }, null, 8, ["modelValue", "header", "default-message"])
+            ])
+          ]),
+          createVNode(unref(script$2))
+        ]),
+        _: 1
+      });
+    };
+  }
+});
+const DesktopUpdateView = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "data-v-8d77828d"]]);
+export {
+  DesktopUpdateView as default
+};
+//# sourceMappingURL=DesktopUpdateView-C-R0415K.js.map

web/assets/DesktopUpdateView-CxchaIvw.css

@@ -0,0 +1,20 @@
+
+.download-bg[data-v-8d77828d]::before {
+  position: absolute;
+  margin: 0px;
+  color: var(--p-text-muted-color);
+  font-family: 'primeicons';
+  top: -2rem;
+  right: 2rem;
+  speak: none;
+  font-style: normal;
+  font-weight: normal;
+  font-variant: normal;
+  text-transform: none;
+  line-height: 1;
+  display: inline-block;
+  -webkit-font-smoothing: antialiased;
+  opacity: 0.02;
+  font-size: min(14rem, 90vw);
+  z-index: 0
+}

web/assets/DownloadGitView-PWqK5ke4.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, l as script, be as useRouter } from "./index-4Hb32CNk.js";
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
+import { d as defineComponent, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, l as script, bi as useRouter } from "./index-Bv0b06LE.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
 const _hoisted_1 = { class: "max-w-screen-sm flex flex-col gap-8 p-8 bg-[url('/assets/images/Git-Logo-White.svg')] bg-no-repeat bg-right-top bg-origin-padding" };
 const _hoisted_2 = { class: "mt-24 text-4xl font-bold text-red-500" };
 const _hoisted_3 = { class: "space-y-4" };
@@ -55,4 +55,4 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
 export {
   _sfc_main as default
 };
-//# sourceMappingURL=DownloadGitView-3STu4yxt.js.map
+//# sourceMappingURL=DownloadGitView-PWqK5ke4.js.map

web/assets/ExtensionPanel-Ba57xrmg.js

@@ -1,8 +1,8 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, U as ref, dl as FilterMatchMode, dr as useExtensionStore, a as useSettingStore, p as onMounted, c as computed, o as openBlock, y as createBlock, z as withCtx, k as createVNode, dm as SearchBox, j as unref, bj as script, m as createBaseVNode, f as createElementBlock, D as renderList, E as toDisplayString, a7 as createTextVNode, F as Fragment, l as script$1, B as createCommentVNode, a4 as script$3, ax as script$4, bn as script$5, dn as _sfc_main$1 } from "./index-4Hb32CNk.js";
-import { g as script$2, h as script$6 } from "./index-nJubvliG.js";
-import "./index-D6zf5KAf.js";
+import { d as defineComponent, T as ref, dx as FilterMatchMode, dC as useExtensionStore, a as useSettingStore, p as onMounted, c as computed, o as openBlock, y as createBlock, z as withCtx, k as createVNode, dy as SearchBox, j as unref, bn as script, m as createBaseVNode, f as createElementBlock, D as renderList, E as toDisplayString, a8 as createTextVNode, F as Fragment, l as script$1, B as createCommentVNode, a5 as script$3, ay as script$4, br as script$5, dz as _sfc_main$1 } from "./index-Bv0b06LE.js";
+import { g as script$2, h as script$6 } from "./index-CgMyWf7n.js";
+import "./index-Dzu9WL4p.js";
 const _hoisted_1 = { class: "flex justify-end" };
 const _sfc_main = /* @__PURE__ */ defineComponent({
   __name: "ExtensionPanel",
@@ -179,4 +179,4 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
 export {
   _sfc_main as default
 };
-//# sourceMappingURL=ExtensionPanel-GE0aOkbr.js.map
+//# sourceMappingURL=ExtensionPanel-Ba57xrmg.js.map

web/assets/GraphView-Bo28XDd0.css

@@ -1,6 +1,5 @@
 
-.comfy-menu-hamburger[data-v-7ed57d1a] {
-    pointer-events: auto;
+.comfy-menu-hamburger[data-v-82120b51] {
     position: fixed;
     z-index: 9999;
     display: flex;
@@ -41,19 +40,19 @@
   z-index: 999;
 }
 
-.p-buttongroup-vertical[data-v-cb8f9a1a] {
+.p-buttongroup-vertical[data-v-27a9500c] {
   display: flex;
   flex-direction: column;
   border-radius: var(--p-button-border-radius);
   overflow: hidden;
   border: 1px solid var(--p-panel-border-color);
 }
-.p-buttongroup-vertical .p-button[data-v-cb8f9a1a] {
+.p-buttongroup-vertical .p-button[data-v-27a9500c] {
   margin: 0;
   border-radius: 0;
 }
 
-.node-tooltip[data-v-46859edf] {
+.node-tooltip[data-v-f03142eb] {
   background: var(--comfy-input-bg);
   border-radius: 5px;
   box-shadow: 0 0 5px rgba(0, 0, 0, 0.4);
@@ -133,13 +132,11 @@
   border-right: 4px solid var(--p-button-text-primary-color);
 }
 
-.side-tool-bar-container[data-v-33cac83a] {
+.side-tool-bar-container[data-v-04875455] {
   display: flex;
   flex-direction: column;
   align-items: center;
 
-  pointer-events: auto;
-
   width: var(--sidebar-width);
   height: 100%;
 
@@ -150,16 +147,16 @@
   --sidebar-width: 4rem;
   --sidebar-icon-size: 1.5rem;
 }
-.side-tool-bar-container.small-sidebar[data-v-33cac83a] {
+.side-tool-bar-container.small-sidebar[data-v-04875455] {
   --sidebar-width: 2.5rem;
   --sidebar-icon-size: 1rem;
 }
-.side-tool-bar-end[data-v-33cac83a] {
+.side-tool-bar-end[data-v-04875455] {
   align-self: flex-end;
   margin-top: auto;
 }
 
-.status-indicator[data-v-8d011a31] {
+.status-indicator[data-v-fd6ae3af] {
   position: absolute;
   font-weight: 700;
   font-size: 1.5rem;
@@ -221,7 +218,7 @@
   border-radius: 0px
 }
 
-[data-v-38831d8e] .workflow-tabs {
+[data-v-6ab68035] .workflow-tabs {
   background-color: var(--comfy-menu-bg);
 }
 
@@ -235,31 +232,36 @@
   border-bottom-right-radius: 0;
 }
 
-.actionbar[data-v-915e5456] {
+.actionbar[data-v-ebd56d51] {
   pointer-events: all;
   position: fixed;
   z-index: 1000;
 }
-.actionbar.is-docked[data-v-915e5456] {
+.actionbar.is-docked[data-v-ebd56d51] {
   position: static;
   border-style: none;
   background-color: transparent;
   padding: 0px;
 }
-.actionbar.is-dragging[data-v-915e5456] {
+.actionbar.is-dragging[data-v-ebd56d51] {
   -webkit-user-select: none;
      -moz-user-select: none;
           user-select: none;
 }
-[data-v-915e5456] .p-panel-content {
+[data-v-ebd56d51] .p-panel-content {
   padding: 0.25rem;
 }
-.is-docked[data-v-915e5456] .p-panel-content {
+.is-docked[data-v-ebd56d51] .p-panel-content {
   padding: 0px;
 }
-[data-v-915e5456] .p-panel-header {
+[data-v-ebd56d51] .p-panel-header {
   display: none;
 }
+.drag-handle[data-v-ebd56d51] {
+  height: -moz-max-content;
+  height: max-content;
+  width: 0.75rem;
+}
 
 .top-menubar[data-v-56df69d2] .p-menubar-item-link svg {
   display: none;
@@ -275,7 +277,7 @@
   border-style: solid;
 }
 
-.comfyui-menu[data-v-929e7543] {
+.comfyui-menu[data-v-68d3b5b9] {
   width: 100vw;
   height: var(--comfy-topbar-height);
   background: var(--comfy-menu-bg);
@@ -288,19 +290,94 @@
   order: 0;
   grid-column: 1/-1;
 }
-.comfyui-menu.dropzone[data-v-929e7543] {
+.comfyui-menu.dropzone[data-v-68d3b5b9] {
   background: var(--p-highlight-background);
 }
-.comfyui-menu.dropzone-active[data-v-929e7543] {
+.comfyui-menu.dropzone-active[data-v-68d3b5b9] {
   background: var(--p-highlight-background-focus);
 }
-[data-v-929e7543] .p-menubar-item-label {
+[data-v-68d3b5b9] .p-menubar-item-label {
   line-height: revert;
 }
-.comfyui-logo[data-v-929e7543] {
+.comfyui-logo[data-v-68d3b5b9] {
   font-size: 1.2em;
   -webkit-user-select: none;
      -moz-user-select: none;
           user-select: none;
   cursor: default;
 }
+
+.comfyui-body[data-v-e89d9273] {
+  grid-template-columns: auto 1fr auto;
+  grid-template-rows: auto 1fr auto;
+}
+
+/**
+  +------------------+------------------+------------------+
+  |                                                        |
+  |  .comfyui-body-                                        |
+  |       top                                              |
+  | (spans all cols)                                       |
+  |                                                        |
+  +------------------+------------------+------------------+
+  |                  |                  |                  |
+  | .comfyui-body-   |   #graph-canvas  | .comfyui-body-   |
+  |      left        |                  |      right       |
+  |                  |                  |                  |
+  |                  |                  |                  |
+  +------------------+------------------+------------------+
+  |                                                        |
+  |  .comfyui-body-                                        |
+  |      bottom                                            |
+  | (spans all cols)                                       |
+  |                                                        |
+  +------------------+------------------+------------------+
+*/
+.comfyui-body-top[data-v-e89d9273] {
+  order: -5;
+  /* Span across all columns */
+  grid-column: 1/-1;
+  /* Position at the first row */
+  grid-row: 1;
+  /* Top menu bar dropdown needs to be above of graph canvas splitter overlay which is z-index: 999 */
+  /* Top menu bar z-index needs to be higher than bottom menu bar z-index as by default
+  pysssss's image feed is located at body-bottom, and it can overlap with the queue button, which
+  is located in body-top. */
+  z-index: 1001;
+  display: flex;
+  flex-direction: column;
+}
+.comfyui-body-left[data-v-e89d9273] {
+  order: -4;
+  /* Position in the first column */
+  grid-column: 1;
+  /* Position below the top element */
+  grid-row: 2;
+  z-index: 10;
+  display: flex;
+}
+.graph-canvas-container[data-v-e89d9273] {
+  width: 100%;
+  height: 100%;
+  order: -3;
+  grid-column: 2;
+  grid-row: 2;
+  position: relative;
+  overflow: hidden;
+}
+.comfyui-body-right[data-v-e89d9273] {
+  order: -2;
+  z-index: 10;
+  grid-column: 3;
+  grid-row: 2;
+}
+.comfyui-body-bottom[data-v-e89d9273] {
+  order: 4;
+  /* Span across all columns */
+  grid-column: 1/-1;
+  grid-row: 3;
+  /* Bottom menu bar dropdown needs to be above of graph canvas splitter overlay which is z-index: 999 */
+  z-index: 1000;
+  display: flex;
+  flex-direction: column;
+}

web/assets/InstallView-DW9xwU_F.js

@@ -1,9 +1,9 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, U as ref, bm as useModel, o as openBlock, f as createElementBlock, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, bn as script, bh as script$1, ar as withModifiers, z as withCtx, ab as script$2, K as useI18n, c as computed, ai as normalizeClass, B as createCommentVNode, a4 as script$3, a7 as createTextVNode, b5 as electronAPI, _ as _export_sfc, p as onMounted, r as resolveDirective, bg as script$4, i as withDirectives, bo as script$5, bp as script$6, l as script$7, y as createBlock, bj as script$8, bq as MigrationItems, w as watchEffect, F as Fragment, D as renderList, br as script$9, bs as mergeModels, bt as ValidationState, Y as normalizeI18nKey, O as watch, bu as checkMirrorReachable, bv as _sfc_main$7, bw as mergeValidationStates, bc as t, a$ as script$a, bx as CUDA_TORCH_URL, by as NIGHTLY_CPU_TORCH_URL, be as useRouter, ag as toRaw } from "./index-4Hb32CNk.js";
-import { s as script$b, a as script$c, b as script$d, c as script$e, d as script$f } from "./index-hkkV7N7e.js";
+import { d as defineComponent, T as ref, bq as useModel, o as openBlock, f as createElementBlock, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, br as script, bl as script$1, as as withModifiers, z as withCtx, ac as script$2, I as useI18n, c as computed, aj as normalizeClass, B as createCommentVNode, a5 as script$3, a8 as createTextVNode, b9 as electronAPI, _ as _export_sfc, p as onMounted, r as resolveDirective, bk as script$4, i as withDirectives, bs as script$5, bt as script$6, l as script$7, y as createBlock, bn as script$8, bu as MigrationItems, w as watchEffect, F as Fragment, D as renderList, bv as script$9, bw as mergeModels, bx as ValidationState, X as normalizeI18nKey, N as watch, by as checkMirrorReachable, bz as _sfc_main$7, bA as isInChina, bB as mergeValidationStates, bg as t, b3 as script$a, bC as CUDA_TORCH_URL, bD as NIGHTLY_CPU_TORCH_URL, bi as useRouter, ah as toRaw } from "./index-Bv0b06LE.js";
+import { s as script$b, a as script$c, b as script$d, c as script$e, d as script$f } from "./index-SeIZOWJp.js";
 import { P as PYTHON_MIRROR, a as PYPI_MIRROR } from "./uvMirrors-B-HKMf6X.js";
-import { _ as _sfc_main$8 } from "./BaseViewTemplate-v6omkdXg.js";
+import { _ as _sfc_main$8 } from "./BaseViewTemplate-BTbuZf5t.js";
 const _hoisted_1$5 = { class: "flex flex-col gap-6 w-[600px]" };
 const _hoisted_2$5 = { class: "flex flex-col gap-4" };
 const _hoisted_3$5 = { class: "text-2xl font-semibold text-neutral-100" };
@@ -314,6 +314,7 @@ const _sfc_main$4 = /* @__PURE__ */ defineComponent({
     const pathExists = ref(false);
     const appData = ref("");
     const appPath = ref("");
+    const inputTouched = ref(false);
     const electron = electronAPI();
     onMounted(async () => {
       const paths = await electron.getSystemPaths();
@@ -355,6 +356,13 @@ const _sfc_main$4 = /* @__PURE__ */ defineComponent({
         pathError.value = t2("install.failedToSelectDirectory");
       }
     }, "browsePath");
+    const onFocus = /* @__PURE__ */ __name(() => {
+      if (!inputTouched.value) {
+        inputTouched.value = true;
+        return;
+      }
+      validatePath(installPath.value);
+    }, "onFocus");
     return (_ctx, _cache) => {
       const _directive_tooltip = resolveDirective("tooltip");
       return openBlock(), createElementBlock("div", _hoisted_1$3, [
@@ -370,10 +378,16 @@ const _sfc_main$4 = /* @__PURE__ */ defineComponent({
                     _cache[0] || (_cache[0] = ($event) => installPath.value = $event),
                     validatePath
                   ],
-                  class: normalizeClass(["w-full", { "p-invalid": pathError.value }])
+                  class: normalizeClass(["w-full", { "p-invalid": pathError.value }]),
+                  onFocus
                 }, null, 8, ["modelValue", "class"]),
                 withDirectives(createVNode(unref(script$5), { class: "pi pi-info-circle" }, null, 512), [
-                  [_directive_tooltip, _ctx.$t("install.installLocationTooltip")]
+                  [
+                    _directive_tooltip,
+                    _ctx.$t("install.installLocationTooltip"),
+                    void 0,
+                    { top: true }
+                  ]
                 ])
               ]),
               _: 1
@@ -595,13 +609,12 @@ const _sfc_main$2 = /* @__PURE__ */ defineComponent({
       }
     });
     return (_ctx, _cache) => {
-      const _component_UrlInput = _sfc_main$7;
       return openBlock(), createElementBlock("div", _hoisted_1$1, [
         createBaseVNode("div", _hoisted_2$1, [
           createBaseVNode("h3", _hoisted_3$1, toDisplayString(_ctx.$t(`settings.${normalizedSettingId.value}.name`)), 1),
           createBaseVNode("p", _hoisted_4$1, toDisplayString(_ctx.$t(`settings.${normalizedSettingId.value}.tooltip`)), 1)
         ]),
-        createVNode(_component_UrlInput, {
+        createVNode(_sfc_main$7, {
           modelValue: modelValue.value,
           "onUpdate:modelValue": _cache[0] || (_cache[0] = ($event) => modelValue.value = $event),
           "validate-url-fn": /* @__PURE__ */ __name((mirror) => unref(checkMirrorReachable)(mirror + (_ctx.item.validationPathSuffix ?? "")), "validate-url-fn"),
@@ -653,11 +666,24 @@ const _sfc_main$1 = /* @__PURE__ */ defineComponent({
           };
       }
     }, "getTorchMirrorItem");
-    const mirrors = computed(() => [
-      [PYTHON_MIRROR, pythonMirror],
-      [PYPI_MIRROR, pypiMirror],
-      [getTorchMirrorItem(__props.device), torchMirror]
-    ]);
+    const userIsInChina = ref(false);
+    onMounted(async () => {
+      userIsInChina.value = await isInChina();
+    });
+    const useFallbackMirror = /* @__PURE__ */ __name((mirror) => ({
+      ...mirror,
+      mirror: mirror.fallbackMirror
+    }), "useFallbackMirror");
+    const mirrors = computed(
+      () => [
+        [PYTHON_MIRROR, pythonMirror],
+        [PYPI_MIRROR, pypiMirror],
+        [getTorchMirrorItem(__props.device), torchMirror]
+      ].map(([item, modelValue]) => [
+        userIsInChina.value ? useFallbackMirror(item) : item,
+        modelValue
+      ])
+    );
     const validationStates = ref(
       mirrors.value.map(() => ValidationState.IDLE)
     );
@@ -942,4 +968,4 @@ const InstallView = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "data-
 export {
   InstallView as default
 };
-//# sourceMappingURL=InstallView-DTDlVr0Z.js.map
+//# sourceMappingURL=InstallView-DW9xwU_F.js.map

web/assets/KeybindingPanel-CDYVPYDp.css

@@ -0,0 +1,8 @@
+
+[data-v-8454e24f] .p-datatable-tbody > tr > td {
+  padding: 0.25rem;
+  min-height: 2rem
+}
+[data-v-8454e24f] .p-datatable-row-selected .actions,[data-v-8454e24f] .p-datatable-selectable-row:hover .actions {
+  visibility: visible
+}

web/assets/KeybindingPanel-DvrUYZ4S.css

@@ -1,8 +0,0 @@
-
-[data-v-2554ab36] .p-datatable-tbody > tr > td {
-  padding: 0.25rem;
-  min-height: 2rem
-}
-[data-v-2554ab36] .p-datatable-row-selected .actions,[data-v-2554ab36] .p-datatable-selectable-row:hover .actions {
-  visibility: visible
-}

web/assets/KeybindingPanel-oavhFdkz.js

@@ -1,9 +1,9 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, c as computed, o as openBlock, f as createElementBlock, F as Fragment, D as renderList, k as createVNode, z as withCtx, a7 as createTextVNode, E as toDisplayString, j as unref, a4 as script, B as createCommentVNode, U as ref, dl as FilterMatchMode, an as useKeybindingStore, L as useCommandStore, K as useI18n, Y as normalizeI18nKey, w as watchEffect, aR as useToast, r as resolveDirective, y as createBlock, dm as SearchBox, m as createBaseVNode, l as script$2, bg as script$4, ar as withModifiers, bj as script$5, ab as script$6, i as withDirectives, dn as _sfc_main$2, dp as KeyComboImpl, dq as KeybindingImpl, _ as _export_sfc } from "./index-4Hb32CNk.js";
-import { g as script$1, h as script$3 } from "./index-nJubvliG.js";
-import { u as useKeybindingService } from "./keybindingService-BTNdTpfl.js";
-import "./index-D6zf5KAf.js";
+import { d as defineComponent, c as computed, o as openBlock, f as createElementBlock, F as Fragment, D as renderList, k as createVNode, z as withCtx, a8 as createTextVNode, E as toDisplayString, j as unref, a5 as script, B as createCommentVNode, T as ref, dx as FilterMatchMode, ao as useKeybindingStore, J as useCommandStore, I as useI18n, X as normalizeI18nKey, w as watchEffect, aV as useToast, r as resolveDirective, y as createBlock, dy as SearchBox, m as createBaseVNode, l as script$2, bk as script$4, as as withModifiers, bn as script$5, ac as script$6, i as withDirectives, dz as _sfc_main$2, dA as KeyComboImpl, dB as KeybindingImpl, _ as _export_sfc } from "./index-Bv0b06LE.js";
+import { g as script$1, h as script$3 } from "./index-CgMyWf7n.js";
+import { u as useKeybindingService } from "./keybindingService-DyjX-nxF.js";
+import "./index-Dzu9WL4p.js";
 const _hoisted_1$1 = {
   key: 0,
   class: "px-2"
@@ -96,6 +96,16 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
     }
     __name(removeKeybinding, "removeKeybinding");
     function captureKeybinding(event) {
+      if (!event.shiftKey && !event.altKey && !event.ctrlKey && !event.metaKey) {
+        switch (event.key) {
+          case "Escape":
+            cancelEdit();
+            return;
+          case "Enter":
+            saveKeybinding();
+            return;
+        }
+      }
       const keyCombo = KeyComboImpl.fromEvent(event);
       newBindingKeyCombo.value = keyCombo;
     }
@@ -151,7 +161,7 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
             value: commandsData.value,
             selection: selectedCommandData.value,
             "onUpdate:selection": _cache[1] || (_cache[1] = ($event) => selectedCommandData.value = $event),
-            "global-filter-fields": ["id"],
+            "global-filter-fields": ["id", "label"],
             filters: filters.value,
             selectionMode: "single",
             stripedRows: "",
@@ -216,7 +226,7 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
             visible: editDialogVisible.value,
             "onUpdate:visible": _cache[2] || (_cache[2] = ($event) => editDialogVisible.value = $event),
             modal: "",
-            header: currentEditingCommand.value?.id,
+            header: currentEditingCommand.value?.label,
             onHide: cancelEdit
           }, {
             footer: withCtx(() => [
@@ -275,8 +285,8 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
     };
   }
 });
-const KeybindingPanel = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "data-v-2554ab36"]]);
+const KeybindingPanel = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "data-v-8454e24f"]]);
 export {
   KeybindingPanel as default
 };
-//# sourceMappingURL=KeybindingPanel-C0Nt6GXU.js.map
+//# sourceMappingURL=KeybindingPanel-oavhFdkz.js.map

web/assets/MaintenanceView-DEJCj8SR.css

@@ -63,10 +63,10 @@
 }
 }
 
-[data-v-74b78f7d] .p-tag {
+[data-v-dd50a7dd] .p-tag {
   --p-tag-gap: 0.375rem;
 }
-.backspan[data-v-74b78f7d]::before {
+.backspan[data-v-dd50a7dd]::before {
   position: absolute;
   margin: 0px;
   color: var(--p-text-muted-color);

web/assets/ManualConfigurationView-DTLyJ3VG.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, K as useI18n, U as ref, p as onMounted, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, a4 as script, a$ as script$1, l as script$2, b5 as electronAPI, _ as _export_sfc } from "./index-4Hb32CNk.js";
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
+import { d as defineComponent, I as useI18n, T as ref, p as onMounted, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, a5 as script, b3 as script$1, l as script$2, b9 as electronAPI, _ as _export_sfc } from "./index-Bv0b06LE.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
 const _hoisted_1 = { class: "comfy-installer grow flex flex-col gap-4 text-neutral-300 max-w-110" };
 const _hoisted_2 = { class: "text-2xl font-semibold text-neutral-100" };
 const _hoisted_3 = { class: "m-1 text-neutral-300" };
@@ -71,4 +71,4 @@ const ManualConfigurationView = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scop
 export {
   ManualConfigurationView as default
 };
-//# sourceMappingURL=ManualConfigurationView-DueOvLuK.js.map
+//# sourceMappingURL=ManualConfigurationView-DTLyJ3VG.js.map

web/assets/MetricsConsentView-C80fk2cl.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
-import { d as defineComponent, aR as useToast, K as useI18n, U as ref, be as useRouter, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, a7 as createTextVNode, k as createVNode, j as unref, bn as script, l as script$1, b5 as electronAPI } from "./index-4Hb32CNk.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
+import { d as defineComponent, aV as useToast, I as useI18n, T as ref, bi as useRouter, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, a8 as createTextVNode, k as createVNode, j as unref, br as script, l as script$1, b9 as electronAPI } from "./index-Bv0b06LE.js";
 const _hoisted_1 = { class: "h-full p-8 2xl:p-16 flex flex-col items-center justify-center" };
 const _hoisted_2 = { class: "bg-neutral-800 rounded-lg shadow-lg p-6 w-full max-w-[600px] flex flex-col gap-6" };
 const _hoisted_3 = { class: "text-3xl font-semibold text-neutral-100" };
@@ -83,4 +83,4 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
 export {
   _sfc_main as default
 };
-//# sourceMappingURL=MetricsConsentView-DTQYUF4Z.js.map
+//# sourceMappingURL=MetricsConsentView-C80fk2cl.js.map

web/assets/NotSupportedView-B78ZVR9Z.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, be as useRouter, r as resolveDirective, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, l as script, i as withDirectives, _ as _export_sfc } from "./index-4Hb32CNk.js";
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
+import { d as defineComponent, bi as useRouter, r as resolveDirective, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, l as script, i as withDirectives, _ as _export_sfc } from "./index-Bv0b06LE.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
 const _imports_0 = "" + new URL("images/sad_girl.png", import.meta.url).href;
 const _hoisted_1 = { class: "sad-container" };
 const _hoisted_2 = { class: "no-drag sad-text flex items-center" };
@@ -83,4 +83,4 @@ const NotSupportedView = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "
 export {
   NotSupportedView as default
 };
-//# sourceMappingURL=NotSupportedView-PDDrAb9U.js.map
+//# sourceMappingURL=NotSupportedView-B78ZVR9Z.js.map

web/assets/ServerConfigPanel-BYrt6wyr.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { o as openBlock, f as createElementBlock, m as createBaseVNode, H as markRaw, d as defineComponent, a as useSettingStore, ae as storeToRefs, O as watch, dy as useCopyToClipboard, K as useI18n, y as createBlock, z as withCtx, j as unref, bj as script, E as toDisplayString, D as renderList, F as Fragment, k as createVNode, l as script$1, B as createCommentVNode, bh as script$2, dz as FormItem, dn as _sfc_main$1, b5 as electronAPI } from "./index-4Hb32CNk.js";
-import { u as useServerConfigStore } from "./serverConfigStore-BYbZcbWj.js";
+import { o as openBlock, f as createElementBlock, m as createBaseVNode, H as markRaw, d as defineComponent, a as useSettingStore, af as storeToRefs, N as watch, dJ as useCopyToClipboard, I as useI18n, y as createBlock, z as withCtx, j as unref, bn as script, E as toDisplayString, D as renderList, F as Fragment, k as createVNode, l as script$1, B as createCommentVNode, bl as script$2, dK as FormItem, dz as _sfc_main$1, b9 as electronAPI } from "./index-Bv0b06LE.js";
+import { u as useServerConfigStore } from "./serverConfigStore-D2Vr0L0h.js";
 const _hoisted_1$1 = {
   viewBox: "0 0 24 24",
   width: "1.2em",
@@ -153,4 +153,4 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
 export {
   _sfc_main as default
 };
-//# sourceMappingURL=ServerConfigPanel-DnGhsuUV.js.map
+//# sourceMappingURL=ServerConfigPanel-BYrt6wyr.js.map

web/assets/ServerStartView-B7TlHxYo.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, K as useI18n, U as ref, bk as ProgressStatus, p as onMounted, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, a7 as createTextVNode, E as toDisplayString, j as unref, f as createElementBlock, B as createCommentVNode, k as createVNode, l as script, i as withDirectives, v as vShow, bl as BaseTerminal, b5 as electronAPI, _ as _export_sfc } from "./index-4Hb32CNk.js";
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
+import { d as defineComponent, I as useI18n, T as ref, bo as ProgressStatus, p as onMounted, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, a8 as createTextVNode, E as toDisplayString, j as unref, f as createElementBlock, B as createCommentVNode, k as createVNode, l as script, i as withDirectives, v as vShow, bp as BaseTerminal, b9 as electronAPI, _ as _export_sfc } from "./index-Bv0b06LE.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
 const _hoisted_1 = { class: "flex flex-col w-full h-full items-center" };
 const _hoisted_2 = { class: "text-2xl font-bold" };
 const _hoisted_3 = { key: 0 };
@@ -93,8 +93,8 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
     };
   }
 });
-const ServerStartView = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "data-v-4140d62b"]]);
+const ServerStartView = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "data-v-e6ba9633"]]);
 export {
   ServerStartView as default
 };
-//# sourceMappingURL=ServerStartView-yzYZ8gms.js.map
+//# sourceMappingURL=ServerStartView-B7TlHxYo.js.map

web/assets/ServerStartView-BZ7uhZHv.css

@@ -1,5 +1,5 @@
 
-[data-v-4140d62b] .xterm-helper-textarea {
+[data-v-e6ba9633] .xterm-helper-textarea {
   /* Hide this as it moves all over when uv is running */
   display: none;
 }

web/assets/UserSelectView-C703HOyO.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, aj as useUserStore, be as useRouter, U as ref, c as computed, p as onMounted, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, bf as withKeys, j as unref, bg as script, bh as script$1, bi as script$2, bj as script$3, a7 as createTextVNode, B as createCommentVNode, l as script$4 } from "./index-4Hb32CNk.js";
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
+import { d as defineComponent, ak as useUserStore, bi as useRouter, T as ref, c as computed, p as onMounted, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, bj as withKeys, j as unref, bk as script, bl as script$1, bm as script$2, bn as script$3, a8 as createTextVNode, B as createCommentVNode, l as script$4 } from "./index-Bv0b06LE.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
 const _hoisted_1 = {
   id: "comfy-user-selection",
   class: "min-w-84 relative rounded-lg bg-[var(--comfy-menu-bg)] p-5 px-10 shadow-lg"
@@ -98,4 +98,4 @@ const _sfc_main = /* @__PURE__ */ defineComponent({
 export {
   _sfc_main as default
 };
-//# sourceMappingURL=UserSelectView-DeJDnrF0.js.map
+//# sourceMappingURL=UserSelectView-C703HOyO.js.map

web/assets/WelcomeView-DIFvbWc2.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { d as defineComponent, be as useRouter, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, l as script, _ as _export_sfc } from "./index-4Hb32CNk.js";
-import { _ as _sfc_main$1 } from "./BaseViewTemplate-v6omkdXg.js";
+import { d as defineComponent, bi as useRouter, o as openBlock, y as createBlock, z as withCtx, m as createBaseVNode, E as toDisplayString, k as createVNode, j as unref, l as script, _ as _export_sfc } from "./index-Bv0b06LE.js";
+import { _ as _sfc_main$1 } from "./BaseViewTemplate-BTbuZf5t.js";
 const _hoisted_1 = { class: "flex flex-col items-center justify-center gap-8 p-8" };
 const _hoisted_2 = { class: "animated-gradient-text text-glow select-none" };
 const _sfc_main = /* @__PURE__ */ defineComponent({
@@ -36,4 +36,4 @@ const WelcomeView = /* @__PURE__ */ _export_sfc(_sfc_main, [["__scopeId", "data-
 export {
   WelcomeView as default
 };
-//# sourceMappingURL=WelcomeView-DkwLdayn.js.map
+//# sourceMappingURL=WelcomeView-DIFvbWc2.js.map

web/assets/index-CBxvvAzM.css

@@ -306,6 +306,7 @@
 .litegraph .dialog .dialog-footer {
   height: 50px;
   padding: 10px;
+  margin: 0;
   border-top: 1px solid #1a1a1a;
 }
 
@@ -442,63 +443,6 @@
   color: black;
 }
 
-.litegraph .subgraph_property {
-  padding: 4px;
-}
-
-.litegraph .subgraph_property:hover {
-  background-color: #333;
-}
-
-.litegraph .subgraph_property.extra {
-  margin-top: 8px;
-}
-
-.litegraph .subgraph_property span.name {
-  font-size: 1.3em;
-  padding-left: 4px;
-}
-
-.litegraph .subgraph_property span.type {
-  opacity: 0.5;
-  margin-right: 20px;
-  padding-left: 4px;
-}
-
-.litegraph .subgraph_property span.label {
-  display: inline-block;
-  width: 60px;
-  padding: 0px 10px;
-}
-
-.litegraph .subgraph_property input {
-  width: 140px;
-  color: #999;
-  background-color: #1a1a1a;
-  border-radius: 4px;
-  border: 0;
-  margin-right: 10px;
-  padding: 4px;
-  padding-left: 10px;
-}
-
-.litegraph .subgraph_property button {
-  background-color: #1c1c1c;
-  color: #aaa;
-  border: 0;
-  border-radius: 2px;
-  padding: 4px 10px;
-  cursor: pointer;
-}
-
-.litegraph .subgraph_property.extra {
-  color: #ccc;
-}
-
-.litegraph .subgraph_property.extra input {
-  background-color: #111;
-}
-
 .litegraph .bullet_icon {
   margin-left: 10px;
   border-radius: 10px;
@@ -661,21 +605,6 @@
 .litegraph .dialog .dialog-content {
   display: block;
 }
-.litegraph .dialog .dialog-content .subgraph_property {
-  padding: 5px;
-}
-.litegraph .dialog .dialog-footer {
-  margin: 0;
-}
-.litegraph .dialog .dialog-footer .subgraph_property {
-  margin-top: 0;
-  display: flex;
-  align-items: center;
-  padding: 5px;
-}
-.litegraph .dialog .dialog-footer .subgraph_property .name {
-  flex: 1;
-}
 .litegraph .graphdialog {
   display: flex;
   align-items: center;
@@ -2110,6 +2039,9 @@
   .-right-4{
     right: -1rem;
   }
+  .bottom-0{
+    bottom: 0px;
+  }
   .bottom-\[10px\]{
     bottom: 10px;
   }
@@ -2119,6 +2051,15 @@
   .left-0{
     left: 0px;
   }
+  .left-1\/2{
+    left: 50%;
+  }
+  .left-12{
+    left: 3rem;
+  }
+  .left-2{
+    left: 0.5rem;
+  }
   .left-\[-350px\]{
     left: -350px;
   }
@@ -2128,6 +2069,9 @@
   .top-0{
     top: 0px;
   }
+  .top-2{
+    top: 0.5rem;
+  }
   .top-\[50px\]{
     top: 50px;
   }
@@ -2137,6 +2081,9 @@
   .z-10{
     z-index: 10;
   }
+  .z-20{
+    z-index: 20;
+  }
   .z-\[1000\]{
     z-index: 1000;
   }
@@ -2196,6 +2143,10 @@
     margin-top: 1rem;
     margin-bottom: 1rem;
   }
+  .my-8{
+    margin-top: 2rem;
+    margin-bottom: 2rem;
+  }
   .mb-2{
     margin-bottom: 0.5rem;
   }
@@ -2286,6 +2237,9 @@
   .h-16{
     height: 4rem;
   }
+  .h-48{
+    height: 12rem;
+  }
   .h-6{
     height: 1.5rem;
   }
@@ -2331,6 +2285,9 @@
   .min-h-screen{
     min-height: 100vh;
   }
+  .w-0{
+    width: 0px;
+  }
   .w-1\/2{
     width: 50%;
   }
@@ -2343,12 +2300,21 @@
   .w-16{
     width: 4rem;
   }
+  .w-24{
+    width: 6rem;
+  }
   .w-28{
     width: 7rem;
   }
+  .w-3{
+    width: 0.75rem;
+  }
   .w-3\/12{
     width: 25%;
   }
+  .w-32{
+    width: 8rem;
+  }
   .w-44{
     width: 11rem;
   }
@@ -2458,6 +2424,9 @@
   .cursor-pointer{
     cursor: pointer;
   }
+  .touch-none{
+    touch-action: none;
+  }
   .select-none{
     -webkit-user-select: none;
        -moz-user-select: none;
@@ -2893,6 +2862,10 @@
     --tw-text-opacity: 1;
     color: rgb(239 68 68 / var(--tw-text-opacity));
   }
+  .text-white{
+    --tw-text-opacity: 1;
+    color: rgb(255 255 255 / var(--tw-text-opacity));
+  }
   .underline{
     text-decoration-line: underline;
   }
@@ -3035,8 +3008,6 @@ body {
   height: 100vh;
   margin: 0;
   overflow: hidden;
-  grid-template-columns: auto 1fr auto;
-  grid-template-rows: auto 1fr auto;
   background: var(--bg-color) var(--bg-img);
   color: var(--fg-color);
   min-height: -webkit-fill-available;
@@ -3046,87 +3017,6 @@ body {
   font-family: Arial, sans-serif;
 }
 
-/**
-  +------------------+------------------+------------------+
-  |                                                        |
-  |  .comfyui-body-                                        |
-  |       top                                              |
-  | (spans all cols)                                       |
-  |                                                        |
-  +------------------+------------------+------------------+
-  |                  |                  |                  |
-  | .comfyui-body-   |   #graph-canvas  | .comfyui-body-   |
-  |      left        |                  |      right       |
-  |                  |                  |                  |
-  |                  |                  |                  |
-  +------------------+------------------+------------------+
-  |                                                        |
-  |  .comfyui-body-                                        |
-  |      bottom                                            |
-  | (spans all cols)                                       |
-  |                                                        |
-  +------------------+------------------+------------------+
-*/
-
-.comfyui-body-top {
-  order: -5;
-  /* Span across all columns */
-  grid-column: 1/-1;
-  /* Position at the first row */
-  grid-row: 1;
-  /* Top menu bar dropdown needs to be above of graph canvas splitter overlay which is z-index: 999 */
-  /* Top menu bar z-index needs to be higher than bottom menu bar z-index as by default
-  pysssss's image feed is located at body-bottom, and it can overlap with the queue button, which
-  is located in body-top. */
-  z-index: 1001;
-  display: flex;
-  flex-direction: column;
-}
-
-.comfyui-body-left {
-  order: -4;
-  /* Position in the first column */
-  grid-column: 1;
-  /* Position below the top element */
-  grid-row: 2;
-  z-index: 10;
-  display: flex;
-}
-
-.graph-canvas-container {
-  width: 100%;
-  height: 100%;
-  order: -3;
-  grid-column: 2;
-  grid-row: 2;
-  position: relative;
-  overflow: hidden;
-}
-
-#graph-canvas {
-  width: 100%;
-  height: 100%;
-  touch-action: none;
-}
-
-.comfyui-body-right {
-  order: -2;
-  z-index: 10;
-  grid-column: 3;
-  grid-row: 2;
-}
-
-.comfyui-body-bottom {
-  order: 4;
-  /* Span across all columns */
-  grid-column: 1/-1;
-  grid-row: 3;
-  /* Bottom menu bar dropdown needs to be above of graph canvas splitter overlay which is z-index: 999 */
-  z-index: 1000;
-  display: flex;
-  flex-direction: column;
-}
-
 .comfy-multiline-input {
   background-color: var(--comfy-input-bg);
   color: var(--input-text);
@@ -3541,84 +3431,6 @@ dialog::backdrop {
   justify-content: center;
 }
 
-#comfy-settings-dialog {
-  padding: 0;
-  width: 41rem;
-}
-
-#comfy-settings-dialog tr > td:first-child {
-  text-align: right;
-}
-
-#comfy-settings-dialog tbody button,
-#comfy-settings-dialog table > button {
-  background-color: var(--bg-color);
-  border: 1px var(--border-color) solid;
-  border-radius: 0;
-  color: var(--input-text);
-  font-size: 1rem;
-  padding: 0.5rem;
-}
-
-#comfy-settings-dialog button:hover {
-  background-color: var(--tr-odd-bg-color);
-}
-
-/* General CSS for tables */
-
-.comfy-table {
-  border-collapse: collapse;
-  color: var(--input-text);
-  font-family: Arial, sans-serif;
-  width: 100%;
-}
-
-.comfy-table caption {
-  position: sticky;
-  top: 0;
-  background-color: var(--bg-color);
-  color: var(--input-text);
-  font-size: 1rem;
-  font-weight: bold;
-  padding: 8px;
-  text-align: center;
-  border-bottom: 1px solid var(--border-color);
-}
-
-.comfy-table caption .comfy-btn {
-  position: absolute;
-  top: -2px;
-  right: 0;
-  bottom: 0;
-  cursor: pointer;
-  border: none;
-  height: 100%;
-  border-radius: 0;
-  aspect-ratio: 1/1;
-  -webkit-user-select: none;
-     -moz-user-select: none;
-          user-select: none;
-  font-size: 20px;
-}
-
-.comfy-table caption .comfy-btn:focus {
-  outline: none;
-}
-
-.comfy-table tr:nth-child(even) {
-  background-color: var(--tr-even-bg-color);
-}
-
-.comfy-table tr:nth-child(odd) {
-  background-color: var(--tr-odd-bg-color);
-}
-
-.comfy-table td,
-.comfy-table th {
-  border: 1px solid var(--border-color);
-  padding: 8px;
-}
-
 /* Context menu */
 
 .litegraph .dialog {
@@ -3718,24 +3530,6 @@ dialog::backdrop {
   will-change: transform;
 }
 
-@media only screen and (max-width: 450px) {
-  #comfy-settings-dialog .comfy-table tbody {
-    display: grid;
-  }
-  #comfy-settings-dialog .comfy-table tr {
-    display: grid;
-  }
-  #comfy-settings-dialog tr > td:first-child {
-    text-align: center;
-    border-bottom: none;
-    padding-bottom: 0;
-  }
-  #comfy-settings-dialog tr > td:not(:first-child) {
-    text-align: center;
-    border-top: none;
-  }
-}
-
 audio.comfy-audio.empty-audio-widget {
   display: none;
 }
@@ -3746,7 +3540,6 @@ audio.comfy-audio.empty-audio-widget {
   left: 0;
   width: 100%;
   height: 100%;
-  pointer-events: none;
 }
 
 /* Set auto complete panel's width as it is not accessible within vue-root */
@@ -3926,7 +3719,7 @@ audio.comfy-audio.empty-audio-widget {
     padding-top: 0px
 }
 
-.prompt-dialog-content[data-v-3df70997] {
+.prompt-dialog-content[data-v-4f1e3bbe] {
   white-space: pre-wrap;
 }
 
@@ -3944,17 +3737,17 @@ audio.comfy-audio.empty-audio-widget {
   margin-bottom: 1rem;
 }
 
-.comfy-error-report[data-v-3faf7785] {
+.comfy-error-report[data-v-e5000be2] {
   display: flex;
   flex-direction: column;
   gap: 1rem;
 }
-.action-container[data-v-3faf7785] {
+.action-container[data-v-e5000be2] {
   display: flex;
   gap: 1rem;
   justify-content: flex-end;
 }
-.wrapper-pre[data-v-3faf7785] {
+.wrapper-pre[data-v-e5000be2] {
   white-space: pre-wrap;
   word-wrap: break-word;
 }
@@ -4023,13 +3816,13 @@ audio.comfy-audio.empty-audio-widget {
   padding: 0px;
 }
 
-.form-input[data-v-1451da7b] .input-slider .p-inputnumber input,
-.form-input[data-v-1451da7b] .input-slider .slider-part {
+.form-input[data-v-a29c257f] .input-slider .p-inputnumber input,
+.form-input[data-v-a29c257f] .input-slider .slider-part {
 
     width: 5rem
 }
-.form-input[data-v-1451da7b] .p-inputtext,
-.form-input[data-v-1451da7b] .p-select {
+.form-input[data-v-a29c257f] .p-inputtext,
+.form-input[data-v-a29c257f] .p-select {
 
     width: 11rem
 }
@@ -4319,26 +4112,26 @@ audio.comfy-audio.empty-audio-widget {
     position: relative;
 }
 
-[data-v-250ab9af] .p-terminal .xterm {
+[data-v-873a313f] .p-terminal .xterm {
   overflow-x: auto;
 }
-[data-v-250ab9af] .p-terminal .xterm-screen {
+[data-v-873a313f] .p-terminal .xterm-screen {
   background-color: black;
   overflow-y: hidden;
 }
 
-[data-v-90a7f075] .p-terminal .xterm {
+[data-v-14fef2e4] .p-terminal .xterm {
   overflow-x: auto;
 }
-[data-v-90a7f075] .p-terminal .xterm-screen {
+[data-v-14fef2e4] .p-terminal .xterm-screen {
   background-color: black;
   overflow-y: hidden;
 }
 
-[data-v-03daf1c8] .p-terminal .xterm {
+[data-v-cf0c7d52] .p-terminal .xterm {
   overflow-x: auto;
 }
-[data-v-03daf1c8] .p-terminal .xterm-screen {
+[data-v-cf0c7d52] .p-terminal .xterm-screen {
   background-color: black;
   overflow-y: hidden;
 }
@@ -4650,28 +4443,28 @@ audio.comfy-audio.empty-audio-widget {
   box-sizing: border-box;
 }
 
-.tree-node[data-v-654109c7] {
+.tree-node[data-v-a945b5a8] {
   width: 100%;
   display: flex;
   align-items: center;
   justify-content: space-between;
 }
-.leaf-count-badge[data-v-654109c7] {
+.leaf-count-badge[data-v-a945b5a8] {
   margin-left: 0.5rem;
 }
-.node-content[data-v-654109c7] {
+.node-content[data-v-a945b5a8] {
   display: flex;
   align-items: center;
   flex-grow: 1;
 }
-.leaf-label[data-v-654109c7] {
+.leaf-label[data-v-a945b5a8] {
   margin-left: 0.5rem;
 }
-[data-v-654109c7] .editable-text span {
+[data-v-a945b5a8] .editable-text span {
   word-break: break-all;
 }
 
-[data-v-976a6d58] .tree-explorer-node-label {
+[data-v-e3a237e6] .tree-explorer-node-label {
   width: 100%;
   display: flex;
   align-items: center;
@@ -4684,10 +4477,10 @@ audio.comfy-audio.empty-audio-widget {
  * By setting the position to relative on the parent and using an absolutely positioned pseudo-element,
  * we can create a visual indicator for the drop target without affecting the layout of other elements.
  */
-[data-v-976a6d58] .p-tree-node-content:has(.tree-folder) {
+[data-v-e3a237e6] .p-tree-node-content:has(.tree-folder) {
   position: relative;
 }
-[data-v-976a6d58] .p-tree-node-content:has(.tree-folder.can-drop)::after {
+[data-v-e3a237e6] .p-tree-node-content:has(.tree-folder.can-drop)::after {
   content: '';
   position: absolute;
   top: 0;
@@ -4790,7 +4583,7 @@ audio.comfy-audio.empty-audio-widget {
   vertical-align: top;
 }
 
-[data-v-0bb2ac55] .pi-fake-spacer {
+[data-v-3be51840] .pi-fake-spacer {
   height: 1px;
   width: 16px;
 }

web/assets/index-CgMyWf7n.js

@@ -1,7 +1,7 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { bA as BaseStyle, bB as script$s, bZ as script$t, o as openBlock, f as createElementBlock, as as mergeProps, m as createBaseVNode, E as toDisplayString, bS as Ripple, r as resolveDirective, i as withDirectives, y as createBlock, C as resolveDynamicComponent, bi as script$u, bK as resolveComponent, ai as normalizeClass, co as createSlots, z as withCtx, aU as script$v, cf as script$w, F as Fragment, D as renderList, a7 as createTextVNode, c9 as setAttribute, cv as normalizeProps, A as renderSlot, B as createCommentVNode, b_ as script$x, ce as equals, cA as script$y, br as script$z, cE as getFirstFocusableElement, c8 as OverlayEventBus, cU as getVNodeProp, cc as resolveFieldData, ds as invokeElementMethod, bP as getAttribute, cV as getNextElementSibling, c3 as getOuterWidth, cW as getPreviousElementSibling, l as script$A, bR as script$B, bU as script$C, bJ as script$E, cd as isNotEmpty, ar as withModifiers, d5 as getOuterHeight, bT as UniqueComponentId, cY as _default, bC as ZIndex, bE as focus, b$ as addStyle, c4 as absolutePosition, c0 as ConnectedOverlayScrollHandler, c1 as isTouchDevice, dt as FilterOperator, bI as script$F, cs as script$G, bH as FocusTrap, k as createVNode, bL as Transition, bf as withKeys, c6 as getIndex, cu as script$H, cX as isClickable, cZ as clearSelection, ca as localeComparator, cn as sort, cG as FilterService, dl as FilterMatchMode, bO as findSingle, cJ as findIndexInList, c5 as find, du as exportCSV, cR as getOffset, c_ as isRTL, dv as getHiddenElementOuterWidth, dw as getHiddenElementOuterHeight, dx as reorderArray, bW as removeClass, bD as addClass, ci as isEmpty, cH as script$I, ck as script$J } from "./index-4Hb32CNk.js";
-import { s as script$D } from "./index-D6zf5KAf.js";
+import { bG as BaseStyle, bH as script$s, bX as script$t, o as openBlock, f as createElementBlock, at as mergeProps, m as createBaseVNode, E as toDisplayString, bO as Ripple, r as resolveDirective, i as withDirectives, y as createBlock, C as resolveDynamicComponent, bm as script$u, bR as resolveComponent, aj as normalizeClass, cp as createSlots, z as withCtx, aY as script$v, cf as script$w, F as Fragment, D as renderList, a8 as createTextVNode, c8 as setAttribute, cx as normalizeProps, A as renderSlot, B as createCommentVNode, bY as script$x, ce as equals, cF as script$y, bv as script$z, cJ as getFirstFocusableElement, c7 as OverlayEventBus, cZ as getVNodeProp, cc as resolveFieldData, dD as invokeElementMethod, bK as getAttribute, c_ as getNextElementSibling, c2 as getOuterWidth, c$ as getPreviousElementSibling, l as script$A, bN as script$B, bQ as script$C, cl as script$E, cd as isNotEmpty, as as withModifiers, da as getOuterHeight, bP as UniqueComponentId, d1 as _default, bZ as ZIndex, bL as focus, b_ as addStyle, c3 as absolutePosition, b$ as ConnectedOverlayScrollHandler, c0 as isTouchDevice, dE as FilterOperator, ca as script$F, ct as script$G, cB as FocusTrap, k as createVNode, bI as Transition, bj as withKeys, c5 as getIndex, cv as script$H, d0 as isClickable, d2 as clearSelection, c9 as localeComparator, co as sort, cL as FilterService, dx as FilterMatchMode, bJ as findSingle, cO as findIndexInList, c4 as find, dF as exportCSV, cW as getOffset, d3 as isRTL, dG as getHiddenElementOuterWidth, dH as getHiddenElementOuterHeight, dI as reorderArray, bT as removeClass, bU as addClass, ci as isEmpty, cM as script$I, ck as script$J } from "./index-Bv0b06LE.js";
+import { s as script$D } from "./index-Dzu9WL4p.js";
 var ColumnStyle = BaseStyle.extend({
   name: "column"
 });
@@ -8787,4 +8787,4 @@ export {
   script as h,
   script$l as s
 };
-//# sourceMappingURL=index-nJubvliG.js.map
+//# sourceMappingURL=index-CgMyWf7n.js.map

web/assets/index-Dzu9WL4p.js

@@ -1,6 +1,6 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { bZ as script$1, o as openBlock, f as createElementBlock, as as mergeProps, m as createBaseVNode } from "./index-4Hb32CNk.js";
+import { bX as script$1, o as openBlock, f as createElementBlock, at as mergeProps, m as createBaseVNode } from "./index-Bv0b06LE.js";
 var script = {
   name: "BarsIcon",
   "extends": script$1
@@ -24,4 +24,4 @@ script.render = render;
 export {
   script as s
 };
-//# sourceMappingURL=index-D6zf5KAf.js.map
+//# sourceMappingURL=index-Dzu9WL4p.js.map

web/assets/index-SeIZOWJp.js

@@ -1,6 +1,6 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value2) => __defProp(target, "name", { value: value2, configurable: true });
-import { bA as BaseStyle, bB as script$6, o as openBlock, f as createElementBlock, as as mergeProps, cJ as findIndexInList, c5 as find, bK as resolveComponent, y as createBlock, C as resolveDynamicComponent, z as withCtx, m as createBaseVNode, E as toDisplayString, A as renderSlot, B as createCommentVNode, ai as normalizeClass, bO as findSingle, F as Fragment, bL as Transition, i as withDirectives, v as vShow, bT as UniqueComponentId } from "./index-4Hb32CNk.js";
+import { bG as BaseStyle, bH as script$6, o as openBlock, f as createElementBlock, at as mergeProps, cO as findIndexInList, c4 as find, bR as resolveComponent, y as createBlock, C as resolveDynamicComponent, z as withCtx, m as createBaseVNode, E as toDisplayString, A as renderSlot, B as createCommentVNode, aj as normalizeClass, bJ as findSingle, F as Fragment, bI as Transition, i as withDirectives, v as vShow, bP as UniqueComponentId } from "./index-Bv0b06LE.js";
 var classes$4 = {
   root: /* @__PURE__ */ __name(function root(_ref) {
     var instance = _ref.instance;
@@ -536,4 +536,4 @@ export {
   script as d,
   script$4 as s
 };
-//# sourceMappingURL=index-hkkV7N7e.js.map
+//# sourceMappingURL=index-SeIZOWJp.js.map

web/assets/keybindingService-DyjX-nxF.js

@@ -1,6 +1,6 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { an as useKeybindingStore, L as useCommandStore, a as useSettingStore, dp as KeyComboImpl, dq as KeybindingImpl } from "./index-4Hb32CNk.js";
+import { ao as useKeybindingStore, J as useCommandStore, a as useSettingStore, dA as KeyComboImpl, dB as KeybindingImpl } from "./index-Bv0b06LE.js";
 const CORE_KEYBINDINGS = [
   {
     combo: {
@@ -186,7 +186,7 @@ const useKeybindingService = /* @__PURE__ */ __name(() => {
       return;
     }
     const target = event.composedPath()[0];
-    if (!keyCombo.hasModifier && (target.tagName === "TEXTAREA" || target.tagName === "INPUT" || target.tagName === "SPAN" && target.classList.contains("property_value"))) {
+    if (keyCombo.isReservedByTextInput && (target.tagName === "TEXTAREA" || target.tagName === "INPUT" || target.tagName === "SPAN" && target.classList.contains("property_value"))) {
       return;
     }
     const keybinding = keybindingStore.getKeybinding(keyCombo);
@@ -247,4 +247,4 @@ const useKeybindingService = /* @__PURE__ */ __name(() => {
 export {
   useKeybindingService as u
 };
-//# sourceMappingURL=keybindingService-BTNdTpfl.js.map
+//# sourceMappingURL=keybindingService-DyjX-nxF.js.map

web/assets/serverConfigStore-D2Vr0L0h.js

@@ -1,6 +1,6 @@
 var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
-import { I as defineStore, U as ref, c as computed } from "./index-4Hb32CNk.js";
+import { a1 as defineStore, T as ref, c as computed } from "./index-Bv0b06LE.js";
 const useServerConfigStore = defineStore("serverConfig", () => {
   const serverConfigById = ref({});
   const serverConfigs = computed(() => {
@@ -87,4 +87,4 @@ const useServerConfigStore = defineStore("serverConfig", () => {
 export {
   useServerConfigStore as u
 };
-//# sourceMappingURL=serverConfigStore-BYbZcbWj.js.map
+//# sourceMappingURL=serverConfigStore-D2Vr0L0h.js.map

web/index.html

@@ -6,8 +6,8 @@
     <meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=no">
     <link rel="stylesheet" type="text/css" href="user.css" />
     <link rel="stylesheet" type="text/css" href="materialdesignicons.min.css" />
-    <script type="module" crossorigin src="./assets/index-4Hb32CNk.js"></script>
-    <link rel="stylesheet" crossorigin href="./assets/index-C1Hb_Yo9.css">
+    <script type="module" crossorigin src="./assets/index-Bv0b06LE.js"></script>
+    <link rel="stylesheet" crossorigin href="./assets/index-CBxvvAzM.css">
   </head>
   <body class="litegraph grid">
     <div id="vue-app"></div>

web/scripts/domWidget.js

@@ -0,0 +1,2 @@
+// Shim for scripts/domWidget.ts
+export const DOMWidgetImpl = window.comfyAPI.domWidget.DOMWidgetImpl;

web/templates/image2image.json

@@ -330,7 +330,7 @@
         "Node name for S&R": "CheckpointLoaderSimple"
       },
       "widgets_values": [
-        "v1-5-pruned-emaonly.safetensors"
+        "v1-5-pruned-emaonly-fp16.safetensors"
       ]
     }
   ],
@@ -440,8 +440,8 @@
   "extra": {},
   "version": 0.4,
   "models": [{
-    "name": "v1-5-pruned-emaonly.safetensors",
-    "url": "https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/resolve/main/v1-5-pruned-emaonly.safetensors?download=true",
+    "name": "v1-5-pruned-emaonly-fp16.safetensors",
+    "url": "https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/resolve/main/v1-5-pruned-emaonly-fp16.safetensors?download=true",
     "directory": "checkpoints"
   }]
 }